Earth Observation is far easier to access than it was a decade ago. Data once handled by a narrow set of agencies and specialist teams now circulate through open archives, cloud platforms, browser tools, and shared analytical environments. Those changes have widened entry, lowered some technical barriers, and made new forms of scrutiny possible. Yet public use still fails for more ordinary administrative reasons. Monitoring programmes lose continuity, workflows never become a stable part of institutional operations, and technically available services sit idle when budgets tighten, procurement stalls, staff move on, or no organisation takes responsibility after release. Availability is only one condition of public use. On its own, it secures very little.

Part of the confusion lies in treating openness as a single condition. A sensing asset produces data. An access regime determines who can use them and on what terms. An operational layer turns them into alerts, maps, and monitoring outputs. Organisational uptake relies on ministries, agencies, NGOs, journalists, or community groups having authority, staff, methods, and routines for repeated use. Some arrangements add a further public-facing layer that keeps information available and inspectable across institutions and publics. Each part breaks down differently, and each draws on its own mix of budgets, contracts, stewardship, and administrative effort. An arrangement can look open on paper and remain thin in practice.

Markets can sustain some parts of this landscape. Firms will pay for bespoke analytics, tasking priority, premium delivery, or sector-specific products when the gains are direct and excludable. Public-facing uses are harder to fund that way. Regulatory oversight, early warning, environmental monitoring, and accountability produce benefits that spill across agencies and sectors, often appearing as avoided harm, better timing, or stronger scrutiny rather than revenue to a single buyer. Those gains are real, but they are difficult to capture through individual transactions. Public procurement and anchor demand therefore shape markets in ways private demand rarely will on its own.

That distinction helps separate cases often grouped together under the heading of openness. Carbon Mapper, MethaneSAT, and FireSat involve monitoring capabilities whose social return is easier to defend than to monetise. NICFI centres on purchased access to imagery already in orbit. SERVIR and Digital Earth Africa show what uptake requires inside institutions and regions. Global Forest Watch serves a different function, keeping shared evidence available across journalists, public agencies, NGOs, and researchers who would otherwise work from more fragmented ground. The economics of openness change at each point.

Methane monitoring is a useful test case because it sits close to regulation and disclosure. Carbon Mapper and MethaneSAT were designed to produce open evidence at a scale relevant to verification, reporting, and enforcement. Both lie somewhere between a classic public mission and a customer-led analytics business. Their intended users extended well beyond paying clients, since the point was to support scrutiny by regulators, watchdogs, researchers, and the wider public.

Carbon Mapper brings together scientific leadership, non-profit coordination, philanthropic capital, public collaboration, and commercial aerospace delivery. Led by the non-profit Carbon Mapper and developed with NASA’s Jet Propulsion Laboratory, Planet Labs, CARB, and other scientific and philanthropic partners, it produces data intended to work at facility scales, where methane evidence can support oversight and regulatory action. That leaves it in an awkward funding position. The outputs are more operational than those of a research mission, while the customer base is less clearly bounded than in a conventional commercial analytics offer. Philanthropy absorbed early risk and helped establish the mission. Public and private partners helped move it towards operation. Procurement has started to carry more of the load, with California supporting the capability through its Satellite Data Purchase Program. The case shows how many sources of support may have to be assembled before this kind of monitoring becomes durable.

MethaneSAT followed a different institutional route. Led by the Environmental Defense Fund with scientific and aerospace partners, and backed by major philanthropic funding, it aimed to strengthen independent methane accountability in oil and gas. Open access was central to that purpose. Oversight depended on use beyond self-reporting and beyond a closed circle of paying users.

When the satellite lost contact in 2025, less than two years after launch, the loss exposed fragility beyond the technical failure itself. A mission built for climate accountability and free public use still left continuity, replacement, and long-term stewardship unresolved once the initial coalition had brought it into operation.

Taken together, Carbon Mapper and MethaneSAT occupy a difficult funding position within methane monitoring. Their outputs are more directly usable for accountability, follow-up investigation, and regulatory screening than broad-coverage public systems, while their economics sit uneasily between public infrastructure and customer-led services. Philanthropic finance recurs here because it can move faster than formal procurement, absorb early uncertainty, and help bring a capability into being before governments decide whether to carry it forward. That helps at the establishment stage. It does not settle who funds persistence, who carries replacement risk, or which public institution is prepared to treat the capability as part of its own responsibilities.

Wildfire detection presents a similar funding problem through a different public function. Earlier detection can change dispatch, evacuation, and containment in the first hours of a fire, when small ignitions may still be controlled and delay is costly. The social return can therefore be large even where no broad paying market forms beneath it. FireSat is being assembled through a hybrid arrangement led by Earth Fire Alliance, with Google Research and Google.org involved in development and funding, and Muon Space responsible for building and operating the FireSat Protoflight mission. That mix reduces the immediate burden on ordinary procurement while drawing on private-sector capacity that already exists.

A public wildfire baseline can coexist with commercial offers built around bespoke coverage, premium analytics, and customer-specific products. Even so, that baseline requires clear protection. Building and operating a programme like FireSat gives private firms flight heritage, operational experience, and credibility that can travel into later contracts. Defence demand, premium services, and other customer priorities may become more attractive as the programme matures. Technical success could create pressure to prioritise some users or services over others unless commitments to public use are protected in the programme design.

Methane and wildfire arrive at the same funding difficulty from different directions. Their public returns are broad, visible, and politically intelligible, while paying demand remains fragmented. Hybrid finance can establish the capability. Its persistence turns on who accepts continuing budget responsibility and how the public obligation is protected once commercial incentives deepen.

Norway’s NICFI programme operates at another point in the chain. It did not finance a new sensor. It used public money to buy access rights to existing commercial imagery and make that imagery available to a wide non-commercial user base. Under a 2020 agreement involving Planet Labs, Airbus, and KSAT, Norway opened monthly high-resolution mosaics across more than forty tropical countries without requiring each government, NGO, Indigenous organisation, or research group to procure imagery separately. Public finance was paying for openness itself.

Norway could move quickly. The imagery was already in orbit. Users gained access that would otherwise have remained expensive, fragmented, or both, and commercial supply entered forest governance on altered terms through a contract that built openness into the arrangement.

Continued access then rested on the durability of that contract, the legal design behind it, and the state’s willingness to keep paying after the public rationale had already been established. The first phase concluded in January 2025, and the follow-on procurement was later cancelled in September 2025 after a legal challenge, making clear the fragility of purchased openness.

Forest governance, Indigenous rights, land use, and transparency in deforestation-linked supply chains already involved disputes over evidence and authority. The imagery did not create those disputes. It changed who could enter them and with what evidential footing. Journalists, NGOs, Indigenous organisations, and public agencies could investigate and contest claims more effectively, though that wider non-commercial use still depended on a procurement mechanism carrying most of the access burden.

Access removes one barrier and leaves another untouched. Ministries and agencies still need documented methods, trained staff, infrastructure, and enough control over the work to adapt it to local decisions and keep it alive after the first funding cycle ends.

SERVIR was built around that harder part of the problem. Through regional hubs across Africa, Asia, and Latin America, it linked NASA data with local technical partners and operational priorities around floods, drought, deforestation, food security, air quality, and related concerns. Its approach built capacity through methods, workflows, and operational processes embedded inside institutions expected to use them repeatedly.

EO becomes dependable inside a ministry, technical agency, or regional risk-management body when it supports recurring decisions and can be carried out by staff as part of ordinary practice. SERVIR has documented cases such as crop-insurance support in Kenya and a forest-fire detection and monitoring system in Nepal, which show how EO workflows can become embedded in public decision processes. They are most useful as examples of repeated institutional use. A dataset can remain available for years without changing administrative practice. A workflow gains firmer footing when it becomes part of a seasonal process, a response protocol, or a standing decision routine. Shared repositories, documented methods, and networked training systems can make these practices more portable across hubs and less reliant on any single team.

Budget hierarchies often treat training, method maintenance, and staff continuity as secondary to hardware, platforms, or visible interfaces. Yet regular use usually rests on those less visible forms of support. SERVIR’s more durable achievements came from embedding methods and decision support within organisations so that the work could continue beyond the first grant cycle.

The withdrawal of U.S. government support in 2025 put that institutional base under strain. Parts of the network continued through regional hubs, university partners, technical teams, and practices already built into local work. Trained staff, shared methods, and routine use gave the programme some capacity to continue. Programmes held together mainly by donor finance become vulnerable when that finance is withdrawn. Those rooted in working relationships and locally held methods have a better chance of surviving the same shock.

Digital Earth Africa moves further towards regional public infrastructure. Open archives, analysis-ready data, cloud processing, notebook environments, interfaces, and metadata are assembled within an operational model intended for regular continental use. Built in part on the Open Data Cube, the programme treats openness as infrastructure and places governance and long-term stewardship increasingly within African institutions.

The difference between a technically hosted service and a regionally stewarded one is institutional as well as technical. A hosted service may reduce friction and still leave operational authority elsewhere. A regionally stewarded arrangement distributes management, technical knowledge, and continuing responsibility across institutions on the continent. The location of those capacities affects who can adapt the infrastructure, who can maintain it, and whose priorities shape its development over time. Its longer-term durability rests on use being embedded in national and regional institutions.

Global Forest Watch serves a different function, making forest data, alerts, and analysis available across public agencies, journalists, NGOs, researchers, communities, and companies working under different mandates and with unequal resources. It brought together open forest-monitoring data, cloud-based analysis, forest-change science, and accessible tools in a form that moved information about forest loss beyond specialist repositories and slow reporting cycles. That reach relied on computation, as well as data, being available beyond specialist institutions.

Wider availability changes how different actors can work with overlapping evidence of forest loss. Journalists can identify stories closer to the event. Public agencies can monitor clearance and fire more regularly. Civil society groups can document violations with greater confidence. Companies face closer scrutiny over supply-chain exposure. These are different forms of action, though they draw more often on shared evidence and common tools.

Global Forest Watch occupies a different place from the earlier cases. It is neither a sensing mission, an access-procurement scheme, nor a regional operational service. It works as a shared evidentiary platform. Keeping such a platform alive requires scientific credibility, technical usability, funding continuity, and support from enough institutions and users that withdrawal by any one actor does not immediately collapse the whole arrangement. That breadth of support has been part of its resilience.

Landsat and Copernicus still represent an older model in which Earth Observation is financed and maintained as public infrastructure through long-term state commitment. Many of the newer cases rest on blended funding arrangements and mixed stewardship: hybrid missions, purchased access, regionally embedded operational work, and platforms that keep evidence in circulation across several institutions.

Commercial firms appear throughout these arrangements as builders, operators, suppliers, and sometimes stewards of crucial parts of the chain. Full public ownership of every layer is unnecessary and often implausible. Governments still have to decide which layer is being funded, what public obligation attaches to it, and who carries continuity when contracts lapse, donors withdraw, or commercial incentives shift. At that point procurement is no longer merely technical instrument. It becomes part of governance.

That governance takes shape in ordinary administrative decisions over access rules, embargoes, user categories, service guarantees, procurement terms, maintenance responsibility, and the treatment of the public baseline during emergencies or commercial expansion. Such decisions determine who bears the cost of interruption, whether services narrow towards favoured users, and whether a capability remains usable once the first funding cycle closes.

Open archives and public platforms have widened access, lowered entry costs, and enabled forms of transparency that once required far more concentrated technical power. Yet release is not the same as provision. Missions, access regimes, operational layers, organisational uptake, and shared evidentiary platforms fail in different ways and require different mixtures of public finance, contracts, stewardship, and institutional support.

Some forms of observation and public use require collective funding even when they do not resemble conventional public works. A state may need to sustain a mission, buy access, underwrite an operational layer, support regional stewards, or create anchor demand where no broad commercial market is likely to emerge. When EO supports environmental oversight, climate governance, disaster preparedness, or fairer access to information, public responsibility has to be built into the arrangement from the start.

Earth Observation now combines abundant technical capacity with a far less settled understanding of how collective-use functions are paid for, governed, and kept in use. New sensors, wider visibility, and easier access will not resolve that administrative burden. The underlying choice is whether societies are willing to sustain functions whose public return is clear even when the revenue case is weak. The economics of openness in Earth Observation do not end when data are released. They begin there. Observation serves a public purpose only when access rights, infrastructure, budgets, operational responsibility, and organisational routines hold together long enough for information to shape repeated decisions, oversight, and scrutiny.

Sources / Further Reading

• Planet Awarded as a Subcontractor for $95M California Satellite Data Purchase Program [link]

• How MethaneSAT will hold polluters accountable [link]

• Methane tracker lost in space [link]

• Inside the launch of FireSat, a system to find wildfires earlier [link]

• NICFI Satellite Data Program – Imagery of the World’s Tropical Forests [link]

• Norway cancels procurement for Tropical Forest Satellite Data Program [link]

• SERVIR – 20 years of Connecting Space to Village [link]

• NASA Data Aid Food Security Assessments in Kenya [link]

• Monitoring the World’s Forests with Global Forest Watch [link]

• Bloomberg Philanthropies Commits $25M to Accelerate Satellite Technologies That Pinpoint Methane Emitters to Turbocharge Fight Against Climate Change [link]

• Planet to Provide Carbon Mapper, Inc. with Hyperspectral Data Until 2030 [link]

• New Zealand to Invest $16 million in MethaneSAT [link]

• Earth Fire Alliance Releases First Wildfire Images from FireSat Protoflight [link]

• Norway’s International Climate and Forest Initiative (NICFI) [link]

• Lessons Learned in the Implementation of a Training Knowledge Management System for the SERVIR Network [link]

• Global Forest Cover Change [link]

• Geoscience Australia appoints the Long-Term Owner of the Digital Earth Africa Program [link]

• Digital Earth Africa 2024 Annual Report [link]

• A Landsat/Copernicus open-data economics source [link]

• The impact of near-real-time deforestation alerts across the tropics [link]

• Landsat’s Economic Value increases to $25.6 Billion in 2023 [link]

Looking back, Earth Observation did not lack progress in 2025. Instead, the year exposed a growing mismatch between a sector that continued to describe itself as broadly commercial and one shaped by defence-led demand.

Those trends had been visible for years: defence alignment, AI everywhere, and commercial consolidation. In 2025, most major announcements revolved around them.

Budgets made priorities explicit

In Europe, Earth Observation was treated unambiguously as strategic infrastructure. The ESA Ministerial Council committed a record €22.3 billion, including €1.2 billion allocated to European Resilience from Space. While environmental monitoring sits at the core of Europe’s EO identity, this programme was framed primarily around surveillance, security, and system resilience, explicitly positioning Earth Observation as a sovereign capability.

Across the Atlantic, the picture was far less stable. The proposed FY2026 U.S. budget introduced deep uncertainty for civilian Earth Observation, with substantial cuts targeting NASA Earth Science, NOAA, and the USGS. Landsat Next, long assumed to be institutionally inevitable, became politically fragile.

This was not a universal retreat. Defence-oriented geospatial procurement in the United States remained aggressively funded. The contrast lay in how Earth Observation outside national security was positioned. Europe treated Earth Observation as a core component of strategic autonomy. The United States appeared undecided about Earth Observation as a civilian public good.

Earth Observation organised around defence contracts

Over the course of 2025, it became difficult to ignore how deeply the sector was entrenching itself into defence and intelligence procurement. This shift was not simply about defence becoming a larger customer, but about defence contracts shaping how systems were designed, funded, evaluated, and prioritised.

As defence procurement logic became more influential, expectations around revisit, latency, automation, and operational responsiveness moved from being use-case specific to baseline design assumptions.

In parallel, purely commercial, non-defence B2B offerings became harder to justify and sustain, not because they lacked value, but because they operated under very different procurement and investment rationales.

Defence demand was no longer simply a growing segment of the market. It had become the reference point around which much of the sector organised itself. According to industry analysts, defence-related demand now accounts for close to half of the Earth Observation market, particularly when data services and value-added analytics are included. Critically, this growth concentrated in precisely the segments that influence constellation design, tasking strategy, and downstream operational workflows.

This orientation was visible across the year’s major contracts: large commercial data awards by the U.S. National Geospatial-Intelligence Agency, long-term sovereign agreements across the Middle East and Asia, and a steady expansion of EO offerings explicitly aligned with national security requirements.

It was also reflected in the year’s most visible commercial announcements. Planet’s Pelican agreement with SKY Perfect JSAT reinforced where high-resolution capacity is being positioned and financed. BlackSky’s Gen-3 contracts and product direction tied high-resolution monitoring closely to automated analytics and operational delivery. Maxar’s reorganisation and rebranding, emphasising sovereign defence capabilities and long-term government relationships, carried the same signal. Individually, none of these moves were surprising. Collectively, they made the market’s centre of gravity harder to deny.

Here, the increasing prominence of “dual-use” framing deserves careful interpretation. In practice, dual-use rarely functioned as a balanced operating model between civil and military objectives. Instead, it served as a framing mechanism, allowing systems and missions to be presented as broadly applicable while being justified, evaluated, and funded primarily through defence-driven requirements.

That distinction matters because it shapes what is funded first. Speed, responsiveness, and actionability were consistently prioritised ahead of calibration, interoperability, and long-term continuity. These latter qualities were not removed from Earth Observation, but they were more often deferred, positioned as secondary concerns to be addressed once operational demands were met.

The practical consequence is that non-defence, non-intelligence B2B Earth Observation now occupies a narrower space. That space still exists, but operating within it requires far more deliberate positioning than before.

Some non-defence segments shifted further downstream, with EO operating as one component inside broader analytical or operational systems rather than as a standalone product, with value realised at the workflow level rather than the data layer.

Launches told the same story from orbit

From a technical standpoint, 2025 was a strong year for Earth Observation.

Institutionally, continuity and precision were central. Copernicus extended its backbone with the launches of Sentinel-1D, Sentinel-5A, and Sentinel-6B, while Sentinel-1C completed commissioning. These missions reinforced long-term measurement priorities and the value of sustained, calibrated observation.

Commercially, however, the emphasis was different. BlackSky’s Gen-3 satellites pushed optical resolution to 35 cm and were tightly integrated with automated analytics. Maxar completed deployment of the WorldView Legion constellation. Planet accelerated its Pelican launches and, notably, flew them with onboard AI compute. Pixxel placed high-resolution hyperspectral capability into orbit with the first satellites of its Fireflies constellation.

These systems are not primarily designed around comprehensive, long-term archiving of the Earth. They are designed to prioritise, filter, and act on observations quickly, in some cases before data ever reaches the ground.

The year also included a reminder of spaceflight’s inherent fragility. The loss of contact with MethaneSAT underscored both the technical difficulty of operating in orbit and the relative vulnerability of climate-first missions when compared with defence-backed programmes.

This contrast reveals how different parts of the sector absorb risk and define what counts as acceptable performance.

Looking ahead, Europe’s next generation of Sentinel missions points toward a higher reference level for institutional Earth Observation. Planned increases in spatial resolution and instrument capability do more than expand what can be observed; they implicitly redefine what is considered acceptable in terms of data quality, consistency, and continuity, particularly as public and commercial datasets are used together within the same analytical and operational workflows.

For commercial providers, this has practical consequences. Resolution alone becomes a weaker differentiator when institutional missions function as quality benchmarks. As spatial detail increases, tolerance for artefacts, misalignment, and inconsistency narrows, and attention shifts toward how data are calibrated, validated, and maintained over time. Higher resolution does not automatically translate into greater confidence; in many cases, it simply makes limitations easier to see.

As resolution increases, institutional datasets also become harder to dismiss as “coarse but free”, and in some domains they begin to function as credible reference layers. This raises the bar for commercial providers, who must increasingly justify not only higher spatial detail, but also calibration quality, consistency, and reliability over time.

AI crossed into operations, while scrutiny thinned

The most consequential AI development in Earth Observation in 2025 was not a startup launch or a product demo. It was the decision by the European Centre for Medium-Range Weather Forecasts to take its Artificial Intelligence Forecasting System (AIFS) into full operational use, running alongside established numerical weather prediction models.

This was not an abrupt shift. AIFS entered active service after years of benchmarking, validation, and institutional review. It was framed as complementary to physics-based models, with parallel execution enabling continuous assessment and trust-building. Reported gains, including improved tropical cyclone tracking, faster inference, and substantially lower energy consumption, were treated as empirical evidence, not marketing claims.

That distinction matters. It demonstrates how AI can be integrated responsibly into Earth Observation when governance, evaluation, and domain expertise are treated as central.

Across much of the commercial EO sector, conditions were different. AI adoption accelerated faster than the frameworks needed to govern its use. Speed, demonstrability, and time-to-market outweighed long-term validation. Foundation models proliferated, natural-language interfaces promised simplified access to geospatial data, and super-resolution approaches were promoted as delivering substantial apparent spatial gains.

In this environment, validation served to support product narratives rather than to challenge them. Evaluation focused on a narrow set of reassuring metrics, while questions of physical limits, robustness, uncertainty, and interoperability were deferred. Systems that produced outputs which looked plausible and could be demonstrated convincingly were often treated as sufficiently validated for use.

This dynamic was reinforced by a rapid influx of machine-learning practitioners into the field. Their contributions accelerated experimentation and development. At the same time, the domain-specific understanding required to interrogate inputs, recognise established patterns, and interpret outputs within physical constraints was not always equally prioritised. Attention concentrated on model performance and visual outputs, while data provenance, consistency, and continuity were postponed.

AI did not fail Earth Observation in 2025. It succeeded well enough, particularly in commercial settings, that scrutiny was treated as optional rather than essential.

This is where a different form of technical debt begins to accumulate, not in code, but in shared understanding, interpretability, and trust.

Better infrastructure, quieter progress

Some of the most consequential progress in 2025 occurred outside headline narratives of AI and defence.

The SpatioTemporal Asset Catalog (STAC) became an official OGC Community Standard. Zarr-based workflows matured. Cloud-native access grew faster and more interoperable. Time series access at scale improved dramatically. From a systems perspective, Earth Observation data became easier to discover, access, and process than at any point before.

These advances were incremental rather than spectacular. They did not produce the kinds of demos or major announcements that attract headlines, but they strengthened the foundations upon which both scientific continuity and commercial reliability depend.

Foundation models, and the return of old mistakes

By 2025, investment in large, generic foundation models for Earth Observation reached unprecedented levels. The question was no longer whether such models could be built, but whether their investment logic made sense outside a narrow set of institutional environments.

Training and maintaining these models requires sustained capital, repeated large-scale iteration, and continuous feedback loops. Even when labour costs are excluded, assuming PhD students or publicly funded researchers, the direct compute cost alone for models comparable in scale to recent flagship efforts can approach €1–2 million, largely in cloud credits. This does not account for downstream fine-tuning, evaluation cycles, or live deployment, nor for the associated energy use and carbon footprint. These costs are real, recurring, and unevenly distributed.

Outside organisations with stable public funding, long-term compute commitments, and mature evaluation pipelines, it is unclear who absorbs these costs, how failure is surfaced, or how iteration is prioritised. In practice, many actors can afford to train once, but far fewer can afford to revise, maintain, and meaningfully govern these models over time.

In this sense, foundation models risk repeating a familiar pattern in Earth Observation: the promise of a single, universal system capable of supporting a wide range of use cases, markets, and revenue streams. Historically, such “one platform that does everything” approaches have struggled. They centralise complexity, push evaluation downstream, and assume a level of user capacity that rarely exists in practice. The result is often a powerful core system that few users are equipped to interrogate, adapt, or meaningfully influence.

There is genuine value in foundation models. Their representation spaces are rich, and they capture complex spatial and temporal structure across large volumes of data. But representation alone does not justify investment. The critical question is whether downstream ecosystems are prepared to use these models responsibly, to evaluate performance against real operational decisions, to integrate feedback from failure cases, and to sustain iteration when models underperform in specific regions or conditions.

This readiness is uneven in structural, rather than technical, ways. In many organisations, benchmarks are still tied to pre-training objectives rather than decision outcomes, and processes for uncertainty characterisation, regional stress-testing, or post-deployment evaluation remain limited. There are teams that do invest in these practices, but they are the exception rather than the norm. As a result, model performance is often assessed using metrics that are convenient to compute and broadly reassuring, while value is claimed in downstream contexts such as risk assessment, monitoring, or planning, where those metrics are only weakly informative.

The focus on global scale reinforces this misalignment. Most Earth Observation users do not buy “global performance.” They pay for coverage, still priced per square kilometre in many cases, and they care about reliability in specific regions, for specific phenomena, under known conditions. Models optimised to maximise average global performance can appear strong overall while underperforming precisely where commercial or policy decisions are made.

In an industry that continues to monetise value primarily per km², scale can become a liability rather than an advantage. The cost of training and maintaining globally optimised models is absorbed upfront, while value is realised locally and unevenly. This creates a structural mismatch between investment logic and how Earth Observation products are priced, sold, and trusted.

Taken together, these dynamics suggest that foundation models are advancing faster than the institutional, economic, and governance structures required to support them responsibly.

Lower barriers, louder noise

AI-assisted development significantly lowered the barrier to producing tools across the Earth Observation ecosystem. This enabled experimentation, learning, and rapid prototyping, and in many cases accelerated individual understanding. It also changed what contribution looked like, and what gained visibility.

There is real value in combining existing tools, libraries, and services. Composition has always been part of effective software development in Earth Observation, and AI-assisted workflows have made it easier to explore ideas, test assumptions, and connect components that were previously difficult to integrate. In many cases, this has genuinely broadened participation and lowered entry costs.

The ease of producing convincing demos outpaced the mechanisms needed to assess durability, relevance, and long-term value. Many tools were built and shared primarily to showcase individual capability or technical fluency. They served their immediate purpose, but rarely moved beyond that stage into operational use.

This matters because Earth Observation depends on a relatively small set of foundational open-source libraries, standards, and data-access tools. These projects handle interoperability, numerical robustness, and long-term compatibility. They underpin both commercial platforms and institutional missions, including systems supported by billion-euro investments but maintained with minimal dedicated funding. When they work well, they tend to disappear into the background.

Lowering the barrier to assembling new workflows and demonstrations did not automatically strengthen this shared foundation on which the sector depends. In practice, it often increased dependence on it without increasing care for its maintenance. Contribution took the form of combining existing components rather than sustaining the libraries, formats, and interfaces that make such composition possible. Output increased, but the condition of the underlying infrastructure became harder to see.

As development practices evolve, the imbalance becomes harder to ignore. The industry is beginning to move away from informal, throwaway prototypes toward systems expected to run in production, interoperate reliably, and withstand audit and review over time. Open source plays a different role here. Specifications can define interfaces and expected behaviour, but they do not ensure numerical correctness, performance, or operational reliability. That work still resides in code, tests, documentation, and long-term stewardship.

The issue is not experimentation, nor the use of AI to accelerate development. It is the absence of conditions that reward maintenance, validation, and long-term ownership at the same pace that they reward visible output. As reliance on shared open-source infrastructure continues to grow, this imbalance accumulates a form of technical and organisational debt that is difficult to measure, but costly to ignore.

Climate stayed central, but not structurally

Climate signals continued to intensify. Data confirmed that 2024 crossed the 1.5 °C threshold, and January 2025 became the warmest January on record.

For an industry built to observe planetary change, these should have been defining reference points. Instead, they functioned as background context. They were noted, acknowledged, and then absorbed into a year dominated by geopolitics and AI.

Earth Observation is better equipped than ever to measure climate change. The tools are more capable, the data more accessible, and the analytical capacity more advanced than at any point before. Yet the alignment between measurement and response has weakened. Acting on climate signals competes with other priorities in a world shaped by geopolitical tension, economic uncertainty, and short-term risk management.

Climate response is not only a scientific challenge; it is also an economic one. In an environment where public budgets are under pressure and procurement favours immediate operational value, sustained investment in climate resilience and mitigation becomes harder to justify alongside other competing priorities.

Climate is central to Earth Observation in principle, but less often in how systems are structured, funded, and justified.

Where this leaves Earth Observation

By the end of 2025, Earth Observation was faster, more operational, and more confident in its capabilities. Budgets expanded. AI systems entered production. Standards stabilised. Constellations matured.

At the same time, the year narrowed what the sector is effectively organised around. Defence increasingly sets the pace. AI compresses complexity into outputs. Visibility is rewarded more readily than understanding. Climate urgency remains present, but no longer consistently shapes structure or strategy.

None of this means that Earth Observation has lost its relevance or potential. The opposite is true; the tools, platforms, and analytical frameworks are stronger than ever. What has become less certain is not capability, but intent; the purposes toward which these systems are directed, and the kinds of outcomes they are ultimately designed to serve.

Acceleration, on its own, is not direction. Neither is scale. Nor automation. Direction emerges from the choices embedded in funding models, evaluation practices, and the kinds of work that are rewarded and sustained over time.

Looking ahead, the question is not whether Earth Observation will continue to advance. It will. The more consequential question is whether the disciplines of stewardship, governance, and institutional memory will advance with it, or whether they will quietly erode.

2026 will not resolve that by default. But it will make the consequences of those choices more visible, and harder for the sector to ignore.

Alongside this article, I am publishing a NotebookLM containing all Spectral Reflectance newsletter issues from 2025. Throughout the year, the newsletter functioned as a running record of industry developments as they emerged, including launches, funding decisions, policy signals, and technical developments.

The notebook is not intended as a comprehensive account of everything that occurred across the sector, and some developments were inevitably missed. Rather, it reflects the material I drew on to form and test the arguments presented here. Readers can explore it directly, query it, and trace how specific observations relate to the broader themes discussed in this piece.

But this time, the image never touches a ground station.

Instead, it veers sideways, into a shoebox-sized rack of GPUs bolted to the inside of another satellite. Within seconds, an onboard model flags a new ignition, cross-checks with wind and fuel maps, and beams down a handful of bytes: coordinates, confidence, predicted spread.

On the fire command center’s screen, the alert looks almost trivial; just one more icon on a map.

Behind it, though, is the question more people in cloud and space are beginning to ask out loud:

Are we really going to build data centers in space?

And if we do… is that an act of climate responsibility, or simply a very expensive way of exporting our problems above the Kármán line?

Earth Observation isn’t the sole force pushing compute off-planet, but it may be the first domain where the case becomes impossible to ignore: too much data, too little bandwidth, too slow a path from pixel to decision. This essay argues that EO is uniquely positioned to become the first scalable proving ground for orbital compute: its bottlenecks are structural, its latency demands are high, and no other domain benefits as much from processing data close to the sensor.

The rest of this essay is about why this “radical” idea is on the table at all, what it could unlock for Earth Observation, and what’s standing squarely in the way.

Why we’re even entertaining such a wild idea

Terrestrial data centers already consume a noticeable slice of global electricity. By 2030, US data centers alone could consume around 9% of the country’s electricity. This rising consumption, fueled by AI training and inference, creates a genuine energy policy problem. What used to be “big IT loads” are now grid-scale industrial assets. The same machines that generate photorealistic cats and power large language models are now competing with homes, factories, and public transit for electricity.

This might sound familiar to anyone tracking the sector. Utilities cannot bring multi-hundred-megawatt capacity online fast enough. Some operators — like Microsoft — have been reported to stockpile GPUs that sit idle simply because they lack the power or the “warm shells” to plug them in. Others quietly wheel in temporary gas generators just to get clusters running because the grid cannot yet carry the load.

Cooling multiplies the issue. Hyperscale facilities consume millions of litres of water a day to keep their processors within comfortable temperatures. In regions already struggling with drought, this is not just unpopular, it is politically explosive. Add in the reality that many of these sites sit on precious land near metropolitan areas, and you start to see why permits are slowing, communities are pushing back, and expansion plans look increasingly like a geopolitical game of Tetris.

Meanwhile, the orbital segment faces its own growing constraints. High-resolution optical constellations, SAR fleets, hyperspectral demonstrators, thermal missions, GNSS reflectometry, atmospheric sounders; they are producing exponentially more data than the Radio Frequency (RF) downlink infrastructure was ever designed to support.

The sensors are scaling exponentially; the downlink is scaling linearly. Petabytes of data are being produced every day and ground stations strain to keep up. Latency from acquisition to usable product stretches into hours and, for some use cases, days. For climate and disaster risk, that lag is not an inconvenience; it’s a failure of design.

Put simply: on Earth, we are running out of room, power, and political patience for more data centers. In space, we are running out of time and bandwidth to move and process the data we already collect.

And somewhere between an overburdened grid below and an overburdened downlink above, one group of people began sketching another option:

What if we just put the data center in orbit?
What if compute lived closer to the sensors, and closer to the only power source that scales indefinitely: sunlight?

The Orbital Solution: A New Geography for the Cloud

Reframing space as an industrial environment rather than a distant frontier makes orbital data centers feel less like science fiction and more like a deliberate reallocation of infrastructure.

The physical argument begins with sunlight. On Earth, solar panels contend with atmosphere, clouds, weather, and the simple fact of night. Their capacity factor, the ratio between what they could produce in theory and what they actually produce over time, hovers around a few tens of percent. But in carefully chosen sun-synchronous orbits, panels can achieve capacity factors exceeding 90%, with stable irradiance and minimal intermittency. A satellite can surf the terminator and remain in near-constant daylight.

It is no surprise that some of the most visible proponents of this vision are the people staring at the scaling curves of AI. Jeff Bezos has spoken openly about gigawatt-scale data centers in space within a few decades. Tom Mueller, who helped architect the early SpaceX engines, has argued that if we want the benefits of compute without burning the planet in the process, a meaningful fraction of heavy compute will eventually need to move off-world.

For most domains, that still feels like a leap. For Earth Observation, which already lives half in orbit and half on the ground, the leap is smaller. We already fly constellations, already generate petabytes overhead, and already treat orbit as part of the data pipeline. Giving these sensors access to compute in orbit becomes less a revolution than an overdue symmetry: sensing up there, thinking up there.

A Revolution in Earth Observation: From Bent Pipe to the In-Orbit Compute Fabric

Earth Observation is already indispensable for managing climate risk, food security, and disaster response. Yet the architecture we still rely on to turn satellite data into decisions is, in many ways, fundamentally 20th-century.

For decades, EO has followed a bent-pipe model: satellites acquire data, store it, and then offload everything to ground stations whenever they pass overhead. From there, the data is transferred into data centers or cloud regions, indexed, calibrated, processed, analysed, and eventually turned into maps, alerts, or time series. At a moderate scale, this works. At the scale of modern and planned constellations, it starts to crack.

Meanwhile, the terrestrial ground station infrastructure, which serves as the traditional backbone for satellite operations, faces immense strain as the volume of space-generated data explodes. These stations are indispensable; without them, a satellite cannot update software, receive commands, or shed the terabytes it accumulates each day. Agencies like the USGS, NASA, and NOAA rely on strategic polar stations, such as Norway’s SvalSat, as these are the only spots where satellites in polar orbits can downlink their data and receive commands on every trip around Earth. However, this dependence has created a critical bottleneck: there are only so many passes per day, only so much RF spectrum to go around, and only so much contact time available for a fleet that is expanding far faster than the ground segment built to serve it. With the number of satellites in Low Earth Orbit (LEO) expected to increase by 190% within the next decade, the RF Ground Station infrastructure is predicted to be strained, affecting data quality and reducing the available bandwidth time for each satellite. This reliance also introduces geopolitical and physical vulnerabilities, as demonstrated by events like cable breaks, which can cause delays of several hours for critical datasets, directly affecting applications like weather forecast and civil protection.

To address this reliance and ground station bottleneck, Orbital Edge Computing (OEC) proposes a different flow, shifting the paradigm from the “bent-pipe” model that has defined EO for half a century. Instead of relying on the ground as the first place where data becomes information, OEC pushes part of the processing pipeline back into space. High-throughput radio and optical links, such as Optical Intersatellite Links (OISLs), ferry raw sensor data sideways; from imagers to orbital compute nodes. There, AI models segment, classify, detect anomalies, estimate physical variables, and decide what is worth keeping. The result is not a continuous flood of scenes to Earth, but a curated stream of maps, alerts, and derived products that are far more actionable and far less burdensome to downlink.

The difference for EO missions is more than marginal. Fires can be detected and confirmed in near-real-time, with predicted spread calculated before the first downlink pass even becomes available. Flooded areas can be mapped from SAR data and compared with building footprints while the orbit is still overhead. Cloud-covered scenes can be discarded on-orbit instead of clogging the ground segment with useless scenes. Crop stress indices can be derived while the constellation is still over the fields, and only the derived products ever leave space. In a fire control center, a civil protection room, or an agricultural advisory service, the user sees the same thing — a dot on a map, a polygon, a time series — but behind it the distance between sensing and decision is shorter and far more efficient.

Crucially, once satellites and orbital data centers can talk to each other, they can also begin to task each other. The jargon for this is “tip-and-cue.” A thermal satellite detecting a suspicious hotspot can tip a higher-resolution optical satellite to capture a follow-up pass. A SAR constellation can cue a high-resolution optical satellite when it sees structural change in a wetland or a suspect vessel in a shipping lane. In a mature orbital compute layer, those tips and cues need not wait for a human operator; they become machine-to-machine negotiations, executed at orbital timescales and governed by policy, priority, and mission logic rather than human reaction speed.

Seen from that angle, orbital data centers are not just remote cloud regions in exotic locations. They are where the EO system begins to behave like a distributed orbital compute layer: sensing locally, integrating globally, and acting — by retasking sensors, escalating alerts, or thinning data streams — without waiting for the ground infrastructure to dictate the tempo of the observation cycle.

The First Wave of Orbital Compute Infrastructure

The early market has arrived: funding rounds, demonstrator missions, feasibility studies, and national strategies are already in motion.

In China, ADA Space is pursuing the most aggressively scaled vision to date. Its Three Body Computing Constellation launched its first 12 satellites in May 2025, each equipped with 100 Gb/s optical links and on-orbit accelerators delivering up to 744 trillion operations per second (TOPS). This “first wave” is designed as the seed of a far larger architecture: a 2,800-satellite distributed supercomputer in LEO. For China, this is more than technical ambition. It is a bid to own a strategic high ground where AI, sensing, and communications converge; a national asset as much as a commercial platform.

In the United States, private enterprise and Big Tech are pursuing parallel “moonshot” approaches. Among the commercial leaders is the US startup Starcloud (formerly Lumen Orbit), which treats space not as a novelty but as the only environment capable of sustaining AI’s power appetite. In November 2025, Starcloud launched Starcloud-1, carrying an NVIDIA H100; a chip nearly two orders of magnitude more powerful than anything previously flown. The mission tests wildfire detection, crop monitoring, and vessel identification models running in orbit. But Starcloud’s long-term vision is even more ambitious: a five-gigawatt orbital hypercluster powered by a solar array four square kilometers across, with second-generation satellites carrying Blackwell-class chips expected as early as 2026.

Axiom Space is taking a different route. Rather than standalone compute nodes, it is integrating orbital data center capability directly into Axiom Station, the commercial successor to the ISS. Axiom has flown both an AWS Snowcone and its AxDCU-1 prototype, and plans to field its first operational ODC nodes in 2025. Partnerships with Red Hat and Kepler suggest a focus on national security workloads, autonomous in-orbit R&D, and a steadily expanding commercial edge computing ecosystem.

The cloud hyperscalers are also entering the field. Google’s Project Suncatcher positions orbit as a “research moonshot” to stretch the scaling laws of machine learning beyond the limits of terrestrial energy and cooling. Instead of NVIDIA hardware, Google intends to fly its own Trillium-generation TPUs, with two prototypes scheduled for launch by 2027 in partnership with Planet.

Across the Atlantic, Europe is framing the challenge through policy; one of sovereignty and climate alignment. The ASCEND feasibility study, led by Thales Alenia Space, argues that orbital compute can support Europe’s net-zero ambitions while strengthening digital independence. The roadmap moves from a 50-kilowatt demonstrator around 2031 to a one-gigawatt facility by mid-century, with strong emphasis on lifecycle emissions, regulatory guardrails, and European control of data processed overhead.

Despite their different ambitions, all of these programs intersect with Earth Observation, which makes EO the most immediate proving ground for in-orbit compute. Whether the driver is national security, AI scaling, or digital sovereignty, EO workloads are the first to benefit from moving processing closer to the source.

Yet turning these ambitions into infrastructure requires more than technical daring. It requires confronting the physical and regulatory realities that no strategy document can gloss over. Markets may be signaling demand, governments may see sovereignty and security advantages, and companies may race to stake orbital territory, but none of that suspends the constraints of heat, radiation, mass, and maintainability. Before orbital compute can mature from a geopolitical aspiration into a reliable layer of critical infrastructure, it must first pass through the hard bottlenecks that space itself imposes.

The Hard Challenges: When Physics Pushes Back

If the story ended there, this essay would be a press release. In reality, orbital data centers collide head-on with constraints defined entirely by physics.

The first is thermal management. Space is indeed cold, but it is also a vacuum. On Earth, data centers lean heavily on convection; fans, chilled water, and heat exchangers move energy away from hot components. In orbit there is no air and no water, only radiation. The Stefan–Boltzmann law sets the fundamental limit on how much heat a surface can radiate at a given temperature and emissivity. To dissipate megawatts of heat without running your radiators at dangerously high temperatures, you need radiating surfaces measured in the thousands of square metres, and you need structure, coolant loops, deployment mechanisms, and careful pointing controls to make them work. The result is that the mass of the cooling system can easily equal or exceed the mass of the compute it is cooling.

Power density suffers as a result. On the ground, with immersion and advanced liquid cooling, designers are pushing 30-100 kilowatts per rack and dreaming of more. In orbit, realistic studies suggest that 10-20 kilowatts per rack will be closer to the practical ceiling. Every extra GPU in a thermal design comes with a penalty in area and mass.

The second hard constraint is radiation. Outside the atmosphere and magnetosphere, electronics are exposed to galactic cosmic rays, solar particle events, and the radiation spikes associated with the South Atlantic Anomaly. These cause long-term radiation damage, measured as total ionizing dose, and also trigger unpredictable single-event effects; bit flips in memory, latch-ups in logic, occasional fatal damage to components. Experiments on the International Space Station and other missions have shown unexpectedly high failure rates among commercial solid-state drives flown with minimal protection. Even purpose-designed chips, like newer AI accelerators, have shown surprising vulnerabilities in their high-bandwidth memory subsystems under test.

Mitigating these effects is possible but never free. Shielding adds mass. Redundancy multiplies cost and increases system complexity. Radiation-hardened parts are often generations behind their terrestrial cousins in performance and can be an order of magnitude more expensive.

Storage is where these constraints intersect most sharply. Spinning disks are too fragile and too heavy to fly. Solid-state drives are more robust mechanically, but their dense flash cells and controllers present large targets for radiation. Aggressive error correction, scrubbing, and replication can mask many faults, but the underlying physics does not go away.

Finally, there is the simple fact of distance. Once a data center is in orbit, you cannot send a technician with a flashlight and a spare drive at three in the morning. Everything — hardware, software, and operations — has to be designed from the start for fault tolerance, autonomous recovery, modular replacement, and eventually, robotic servicing. And everything that is not deorbited responsibly becomes part of a growing cloud of debris that threatens the very constellations that orbital compute is meant to serve.

These are not abstract worries; they define the physical boundaries of what can be built. Within those boundaries, though, there is still room for careful engineering and meaningful progress.

The Easier Challenges and the Emerging Enablers

Not all obstacles are laws of nature. Some are matters of engineering maturity and here the picture is more optimistic.

Virtualization is a good example. On Earth, we treat hypervisors and virtual machines as tools for flexibility and utilization. In orbit, they become tools for survival. By inserting a hypervisor such as Xen between the hardware and the operating systems, engineers can carve the processor into strongly isolated domains. Flight software and platform control sit in one or more real-time VMs, pinned to dedicated cores with strict guarantees. EO workloads — fire detection, ship tracking, change detection, even tenant-specific models — run in other VMs that can be restarted, throttled, or migrated without endangering the satellite’s basic functions.

Experiments on representative multi-core SoCs have shown that this kind of configuration can deliver both low-latency, low-jitter behavior for critical tasks and robust containment of badly behaved or heavily loaded workloads. In practical terms, that means:

A runaway ML experiment cannot starve the altitude control loop. A third-party EO tenant cannot crash the mission by accident. A container processing flood extents cannot quietly monopolise resources needed for station-keeping. The result is an orbital compute node that behaves less like a monolithic system where everything shares the same failure points and more like a small in-orbit cluster, where mixed-criticality workloads run side by side without the risk of failing together.

Communications technology is following a similar arc. Optical inter-satellite links, once exotic, are becoming standard in large constellations. For EO and orbital compute, they form the high-bandwidth backbone of the system. They allow imaging satellites to dump data to compute nodes that are not directly overhead a ground station. They allow those nodes to share intermediate products and models among themselves. They carry tip-and-cue messages — machine-to-machine tasking instructions — connecting sensors and compute into a dynamic, reconfigurable sensor-web.

Launch costs and on-orbit servicing sit in a fuzzier but encouraging category. Reusable rockets have already driven the price per kilogram to orbit down dramatically compared to the shuttle era. Fully reusable systems, if they reach operational maturity, may push costs below the critical thresholds where a 10-year total cost of ownership for orbital compute starts to compete with the energy and infrastructure cost of high-end terrestrial facilities. Robotic servicers are being tested that can refuel, reposition, and eventually repair or upgrade satellites. Taken together, these developments do not make orbital data centers cheap, but they make them less absurd.

None of this removes the underlying physical limits. It simply reframes the problem: from “infeasible” to “viable for certain workloads”, provided the design respects those constraints.

Uncertainties in the High Ground: Law, Governance, and Control

Even if engineers manage to solve or at least squeeze these technical problems, there is still the quieter question of how an orbital cloud fits into our legal and political frameworks.

Space law as it stands was not written with data centers in mind. States remain responsible for objects they launch or authorise. Liability for damage is written into treaties. But where does data jurisdiction live when the processors are in orbit? If an EO company based in one country processes imagery of critical infrastructure from another country on a platform built and launched by a third, whose laws apply — and who has the right to inspect, regulate, or restrict it? If tip-and-cue chains autonomously retask satellites over a sensitive region, who is accountable for the action: the operator, the algorithm, the state that licensed the platform, or the tenant whose model initiated the cue?

Regulators are only beginning to grapple with terrestrial questions of data residency and sovereignty. Adding an orbital layer will not simplify that. For EO, which already sits in the sensitive space between transparency and surveillance, the arrival of orbital compute and autonomous machine-to-machine tasking will sharpen old tensions and create new ones.

Environmental governance is another open front. If lifecycle emissions from launch and re-entry turn out to be significant contributors to warming or ozone depletion, it is hard to imagine that large orbital compute constellations will escape scrutiny. One can easily foresee permitting regimes that require full lifecycle climate accounting, mitigation plans, or even offsets for launch and re-entry footprints; mirroring, at altitude, the environmental reviews that already shape terrestrial infrastructure projects.

And then there is debris. Large, long-lived structures are not trivial additions to crowded orbital shells. If multiple states and corporations deploy competing orbital compute hubs without robust end-of-life plans, the probability of collisions and cascading fragmentation grows. For EO, which depends on the long-term stability of low-Earth orbit, this isn’t a distant or theoretical issue but a hard operational constraint on the entire sector.

How This Future Could Look — If We Get It (Mostly) Right

Return, finally, to that fire in the Mediterranean and roll the clock forward to a plausible late-2050s.

Launch costs have fallen into the tens of dollars per kilogram. Fully reusable rockets fly often enough that adding orbital infrastructure is no longer an extraordinary undertaking but simply another predictable operational expense. On-orbit servicing has matured to the point where satellites can be refuelled, nudged into new orbits, and have modular compute “bricks” swapped out robotically. Debris rules are enforced; operators who fail to comply lose access to launch services, insurance, or the radio spectrum needed to run their satellites.

Above us, in a handful of carefully managed orbital shells, sit clusters of orbital EO compute hubs. They are not absurd mega-rings, but more modest stations: trusses, solar arrays, large radiator panels, instrument modules, and both pressurised and unpressurised racks. Each hub carries a layered stack of workloads with different criticalities. Some cores run radiation-hardened (rad-hard) control loops; others host higher-performance accelerators that are expected to live a few years before being swapped. Hypervisors carve the hardware into slices. Spacecraft control lives in one set of virtual machines, infrastructure for communications in another, high-priority EO applications in a third, and time-boxed experiment slots in a fourth.

Around these hubs, constellations of EO satellites orbit in different planes and inclinations. Optical imagers, SAR platforms, thermal sensors, GNSS reflectometry missions, atmospheric sounders, and niche scientific instruments each play their part. Some feed data directly to Earth. Many send their raw output sideways to the hubs, where it is combined with data from other sensors, from reanalysis products, from climate models, and from historical archives cached in-orbit.

Tip-and-cue has become routine. A SAR satellite noticing unusual activity in a coastal zone cues a very-high-resolution optical satellite to take a closer look on its next pass. A thermal mission detecting unusual urban hotspots at night nudges a hyperspectral sensor to sample the area for signs of industrial flaring or leaks. A flood detection model running on a hub identifies a rapidly expanding inundation area and pushes updated tasking patterns to upstream imagers to maximise coverage over the next few orbits. Much of this negotiation happens machine-to-machine, mediated by policies set on the ground but executed at orbital speed.

On Earth, the interfaces look almost disappointingly simple. Fire agencies subscribe to ignition alerts for predefined areas of interest rather than to “all imagery over this region.” River basin managers subscribe to floodplain exceedances and soil moisture anomalies instead of raw scenes. Agricultural advisory services receive updated, field-scale crop stress maps driven by models that were partially computed on-orbit and partially refined on the ground. National climate centers pull continuous, quality-controlled EO-derived indicators without seeing the terabytes of data from which they were derived.

None of this makes terrestrial data centers obsolete. Heavy model training, long-term archival, open data services, exploratory science, offline reanalysis; all of that still happens in large facilities firmly bolted to the crust. The hybrid model emerges naturally: orbit for proximity to sensors and to sunlight; Earth for depth, breadth, and the messy business of humans interrogating data.

In that world, the question is no longer whether we should put data centers in space in some abstract sense, but rather which computations genuinely belong there. The answer will not be “everything” and it will certainly not be “nothing.” It will be a moving boundary, shaped by launch costs, by policy, by climate accounting, and by the unforgiving algebra of power, bandwidth, and latency.

What seems clear already is that EO will be one of the first domains to cross that boundary in a meaningful way. The mismatch between how quickly we can see the planet change and how slowly we can respond is not going away on its own. Orbital compute, paired with autonomous tip-and-cue, strong virtualization, and honest environmental accounting, offers one of the few credible paths to closing that gap.

Used well, it can help transform an overloaded EO ecosystem into an orbital fabric that observes, interprets, and responds on operational timelines.

Used badly, it will become just another shiny, expensive distraction orbiting above a planet that never managed to align its infrastructure with its ecological and physical limits.

Ultimately, technology alone will not decide which of those futures we get. The choices we make now — about where computation happens, how we account for its costs, and which problems deserve to be addressed in orbit — will.

Further Reading & Sources

In researching this piece, I worked through several papers, feasibility studies, mission docs, technical reports, and public statements. I used NotebookLM to organise, query, and cross-reference my sources. You can explore the notebook here.

Sentinel-2C launched into orbit on 5 September at 03:50 CEST, marking a significant milestone for the Copernicus program. In the lead-up to this launch, there has been considerable discussion within the Earth observation community, even sparking a petition advocating for the continued operation of Sentinel-2A.

This satellite, which has outperformed expectations, now faces potential decommissioning as Sentinel-2C prepares to assume its duties.

This raises a critical question:

Should the successful launch of Sentinel-2C automatically lead to the decommissioning of a fully functional and still valuable Sentinel-2A?

Follow the discussion taking place on LinkedIn here and here.

Edit: As Mark Drinkwater, Head of the Earth & Mission Science Division at the European Space Agency (ESA), pointed on the LinkedIn discussion, figures fpr operational costs are listed here.
"The costs for the construction of Sentinel-2C amount to a total of 200 million euros. The annual operating costs are around 25 million euros."

The Case for Extending Sentinel-2A's Mission

One compelling argument for retaining Sentinel-2A in orbit is the enhanced capability of a constellation with three active Sentinel-2 satellites. This configuration would significantly increase revisit frequency, improving the temporal resolution of the data.

Upon launch, Sentinel-2C will enter a 3-6 month commissioning phase, during which its instruments will be calibrated and validated. Once fully operational, it is intended to replace Sentinel-2A, taking over its position in the constellation alongside Sentinel-2B.

Why Keep Sentinel-2A Operational?

Sentinel-2 satellites are designed for an initial mission life of seven years, with the potential to extend operations up to 12 years, depending on fuel reserves. These reserves include fuel designated for a controlled deorbit at the end of the satellite’s mission.

Sentinel-2A, launched in 2015, is therefore theoretically capable of functioning until mid-2027. However, the exact status of its fuel reserves is not publicly disclosed, adding a layer of uncertainty to predictions about its operational longevity. The satellite might require deorbiting earlier than 2027 if the remaining fuel is insufficient to continue operations safely.

One factor influencing the decision to decommission could be the cost of maintaining the satellite. While the exact annual operational costs for Sentinel-2A are not public, they are likely substantial. The petition estimates that these costs could range from several million to a few dozen million euros per year.

Despite these expenses, the petition argues, continuing operations of an existing satellite is generally more economical than building and launching a new one. For instance, Sentinel-2B, which is similar in design, cost several hundred million euros to develop and launch.

Is Retirement the Best Option?

The decision to retire Sentinel-2A is likely driven by a combination of factors, including not only operational costs but also strategic considerations. Extending Sentinel-2A's mission by a mere 1.5 to 2 years might not justify the expense and effort of supporting three satellites in the constellation—unless the benefits of enhanced data collection significantly outweigh these costs.

The Landsat 5 Comparison: A Useful Benchmark?

While some point to the extraordinary 29-year operational life of Landsat 5 as a precedent, it’s important to recognize the differences in mission design and requirements. Landsat 5, launched in 1984, was initially planned as a 3-year mission but far exceeded its expected lifespan due to strategic operational adjustments that conserved fuel and extended its life.

Sentinel-2 is designed to revisit the same location on Earth every 5 days, often in tandem with its twin, Sentinel-2B. To achieve this frequent revisit rate, Sentinel-2A performs regular attitude control maneuvers to ensure its instruments remain precisely aligned with their targets. It also maintains a highly stable orbit to guarantee consistent image quality and calibration throughout its mission.

In contrast, Landsat 5 prioritized long-term data continuity over frequent, precisely-timed revisits. Consequently, it required fewer attitude and orbit adjustments, as its mission did not necessitate the same level of precision as Sentinel-2A.

Moreover, as Landsat 5 aged, NASA and the USGS adapted their operations to conserve fuel, reducing the frequency of certain maneuvers to prolong its operational life.

Finally, it's worth noting that while Sentinel-2 satellites are equipped with specific fuel reserves for a planned end-of-life deorbit maneuver, Landsat 5 was not designed with a deorbit plan, allowing it to continue operations for as long as possible.

Reimagining Sentinel-2A’s Role: A New Mission

One possible avenue for extending Sentinel-2A's mission could involve redefining its operational objectives. If ESA were to adjust its mission profile—perhaps by reducing the frequency of maneuvers and focusing on less frequent, long-term data collection—Sentinel-2A could continue to deliver valuable data for several more years. This approach, similar to the later years of Landsat 5, would conserve fuel and potentially allow Sentinel-2A to operate beyond its current projected end-of-life.

However, such a shift would require careful consideration of the trade-offs involved, particularly regarding data precision and the overall objectives of the Copernicus program.

Conclusion: A Strategic Decision for the Future

The fate of Sentinel-2A is not just a matter of operational capability or cost—it’s a strategic decision that will impact the Copernicus program's ability to provide critical Earth Observation data. Whether to decommission or extend the mission of Sentinel-2A involves balancing the immediate benefits of enhanced data collection against the long-term sustainability of the program.

As Sentinel-2C begins its mission, it's crucial for ESA and the broader Earth observation community to carefully weigh the benefits of extending Sentinel-2A's mission against the program's broader goals.

Follow the discussion taking place on LinkedIn here and here.

References:

• [Petition] Copernicus should keep S2A operational after S2C launch [link]

• Olivier Hagolle's post on LinkedIn [link]

• 's post on LinkedIn [link]

• The Final Journey of Landsat 5: A Decommissioning Story [link]

  ---

One of the major challenges when building applications using EO data is having “clean-ready-to-use-data”, or what we most commonly refer to as Analysis Ready Data (ARD); that is, data that we can use almost out-of-the-box and do time-series analysis, train machine learning models and derive high-level products without spending huge amounts of time in data preparation in order to work on the actual task.

NOTE 1: Although there are efforts towards non-optical ARD (e.g. Sentinel-1 ARD¹), this post covers optical ARD.
NOTE 2: I have split this article into two sections. 
In the first section I explore what we “traditionally” consider as ARD and ARD products such as the Landsat ARD and the Harmonized Landsat and Sentinel-2 (HLS) ARD. I follow ARD-related workshops and highlight several points I find interesting in the presentations. I also dedicate a section related to Planetscope data and Planet’s efforts towards creating an ARD product. 
In the second section, I review a recent study conducted by the Landsat Advisory Group exploring earth observation products, including data, algorithms, and workflows. I dedicate Section 2 of this article on it, as I consider it an “expanded” version of ARD.

Section 1

Defining ARD

There are several definitions of ARD:

• “Data that have been processed to allow analysis with a minimum of additional user effort are often referred to as Analysis Ready Data (ARD).”²

• “(CEOS) Analysis Ready Data (CEOS-ARD) are satellite data that have been processed to a minimum set of requirements and organized into a form that allows immediate analysis with a minimum of additional user effort and interoperability both through time and with other datasets.”³

• products that enable the scientific community and non-expert data users to spend more time conducting research into urgent issues in earth system science, such as climate change, food-energy-water nexus, and other priorities, rather than data cleaning and standardisation procedures³

The last one is not a definitation but rather the goal of ARD products. However, there is something common in all of them; the need for reducing usage complexity.

Why we want (“traditional”) ARD?

Interoperability and Data Quality

Ideally, users would be able to examine any pixel over time seamlessly, in a “stackable” data cube.

ARD provides consistent formats, projections, and coordinate systems across different satellite datasets, enabling seamless integration and analysis.
More than that, ARD is usually accompanied with QA assets; quality assessment information about the data (QA bands for cloud, haze, snow, flagged pixels due to poor quality, etc).

Data compatibility for time-series analysis

ARD facilitates time-series analysis by providing consistent and compatible data across different satellite sensors and time periods. It enables monitoring and comparison of environmental changes, land cover dynamics, and other Earth system processes over extended periods, helping to understand trends and patterns in a consistent manner.

Time Efficiency and Accessibility

ARD reduces the time and effort required for data preprocessing and preparation. By providing analysis-ready data, it eliminates the need for users to perform repetitive and resource-intensive preprocessing steps, such as radiometric and geometric corrections.

“ARD is a function of context”

ARD means something different to different users; in other words, it depends on the use-case⁴.

For example, a researcher who wants to apply their novel method for atmospheric correction, certainly wouldn’t access a L2-product that has already been corrected.

On the other hand, if the task is detecting vegetation change, then users would need a product they could use out-of-the-box (geometrically/radiometrically/atmospherically corrected, provided with clouds masks, etc).

Landsat ARD — The first ARD product

In 2017 the United States Geological Survey (USGS) Earth Resources Observation and Science (EROS) Center made available Landsat ARD for the conterminous United States (CONUS), Alaska, and Hawaii.² 
They set the foundations of what ARD should look like. This was a massive undertaking as USGS processed decades of Landsat data.

These are the points on what ARD products should be, according to [2]:

• tiled

• georegistered

• top of atmosphere (TOA) and atmospherically corrected

• defined in a common equal area projection

• accompanied by spatially explicit quality assessment information (QA bands)

• have non-target features (clouds) and poor quality observations flagged

• accompanied by appropriate metadata to enable further processing while retaining traceability of data provenance

• geometrically and radiometrically consistent

• processed in a community endorsed manner

The USGS had recognized early on that different users have different expectations from an ARD product; some might require atmospherically corrected data, while others may wish to apply their own atmospheric correction methods. So they provided both!

Also, they provide appropriate metadata, such as view and azimuth angles, so that users have all the parameters they need to further process the data (e.g. when deriving a Nadir BRDF-Adjusted Reflectance (NBAR) product).

I encourage you to read the paper [2]. The authors describe the steps they took and the challenges they faced in order to release Landsat ARD Collection 1. (the archive has been reprocessed to Collection 2 in 2020).

A prime feature of the Landsat ARD is its georegistration consistency and the main motivation for re-processing Collection 1 was to improve the Landsat absolute geolocation accuracy even further, using the Sentinel-2 Global Reference Image (GRI). 
This improved the interoperability of the global Landsat archive spatially and temporally.

Achieving high geolocation enables the reliable information extraction from time series (that is, the same pixel corresponds to the same area through time and is not shifted due to misalignment issues).
This feature makes Landsat Collection 2 an “interoperable” product; set of data, originating from different sensors (Landsat 1–9), that can be stacked into a spatiotemproal data cube.

However, Landsat Collection 2 is not a harmonized product (across different Landsat missions).

Harmonized Landsat and Sentinel-2 (HLS)

“The Harmonized Landsat and Sentinel-2 (HLS) project is a NASA initiative aiming to produce a Virtual Constellation (VC) of surface reflectance (SR) data acquired by the Operational Land Imager (OLI) and MultiSpectral Instrument (MSI) aboard Landsat 8 and Sentinel-2 remote sensing satellites, respectively.”⁵

The HLS product merges observations of the land surface from both sensors into a single comprehensive dataset, leveraging their similar spectral, spatial, and angular characteristics, as well as their sun-synchronous orbits.

As the HLS team stated, by “harmonized,” they mean that the products are:

• gridded to a common pixel resolution, map projection, and spatial extent

• atmospherically corrected and cloud-masked to surface reflectance using a common radiative transfer algorithm,

• normalised to a common nadir view geometry through Bi-directional Reflectance Distribution Function (BRDF) estimation, and

• adjusted to represent the response from common spectral bandpasses.

In essence, these harmonized products serve as the foundational “data cube,” allowing users to examine any pixel over time and treat the near-daily reflectance time series as if they originated from a single sensor; this is a main characteristic of harmonisation. 
(More on the differences between interoperability and harmonisation later on)

The potential of the HLS data set lies in its ability to support various applications requiring high temporal and spatial resolutions, with crop monitoring being one of the main drivers.

ARD Workshops

ARD has been such a ̶p̶a̶i̶n̶ challenge, that people organise workshops about it; Namely, the ARD Satellite Data Interoperability Workshop. Also, ARD-related sessions and talks have been taking place in the Joint Agency Commercial Imagery Evaluation (JACIE) and the Very High-resolution Radar & Optical Data Assessment (VH-RODA) Workshops.

Going through the presentations of this year’s workshops (and other recent presentations), I have highlighted some points I have found interesting:

1. Simon Oliver made the case for an “ARD Measurement Quality Goal” through an intercomparison study of data sources for the Digital Earth Australia program. The comparative cases presented:

• Spectral difference between Landsat 8 OLI and Sentinel-2 MSI

• Intercomparison between USGS Landsat 8 SR product with ESA Sentinel-2 SR

• Inter-comparison between Geoscience Australia (GA) Landsat 8 and Sentinel-2 SR products

• Compare multi-sensor time-series of spectral indices NDVI, NDWI, NBR generated from multiple ARD sources over typical Australian sites

2. “ARD is a function of context”
Scott Arko from Descartes Labs had this line in their presentation, which I have used multiple times in this post; I believe it’s a core concept in ARD. 
As already mentioned, ARD depends on the use case.

3. Sen2Like: a processor for harmonising and fusing optical EO data imagery | Enrico Giuseppe Cadau, Serco for ESA, et al. | 2023 ARD Satellite Data Interoperability Workshop

“The Sen2Like demonstration processor has been developed by ESA in the framework of the EU Copernicus programme.
The main goal of Sen2Like is to generate Sentinel-2 like harmonized/fused surface reflectances with higher periodicity by integrating additional compatible optical mission sensors.
It is a contribution to on going worldwide initiatives (*NASA-HLS, FORCE⁶, CESBIO) undertook to facilitate higher level processing starting from harmonized data.”⁷

Sen2Like performs the following steps, depening on the input product type:

• Image stitching of the different tiles

• Geometric corrections including the co-registration to a reference image

• Inter-calibration (for S2 L1C)

• Atmospheric corrections

• Transformation to Nadir Bidirectional Reflectance Distribution Function (BRDF) Adjusted Reflectance (NBAR)

• Application of Spectral Band Adjustment Factor (SBAF) (for LS8/9)

• Production of LS8/LS9 high resolution 10 m pixel spacing data (data fusion)

Sen2Like generates the two following ARD product types:

• Level-2H: harmonized Sentinel-2 and Landsat 8/9 products (native spatial resolutions kept)

• Level-2F: fused Landsat 8/9 to Sentinel-2 products (S2 spatial resolution for Landsat data)

Furthermore, the team has performed studies on the integration of hyperspectral data (DESIS, PRISMA) into Sen2Like! They have proven that the Sen2Like framework is able to process the products of any other multispectral or hyperspectral optical data at the L2A level.⁸

4. “Sometimes ARD is challenging to integrate and not a good fit”

Floodbase has a really good argument for not using ARD; “Disaster response requires data ASAP”.
This isn’t exactly ARD-related, but rather using Data-As-Ready-As-It-Gets. 
However, one of Floodbase’s concerns is the ability to trace errors and know whether a product is pulled and replaced! To their use-case, traceability on this level matters!

5. Greg Stensaas, USGS EROS Cal/Val Center of Excellence (ECCOE) on “The Importance of Calibration and Validation in a Changing World”
Q. “How do you get thousands of Earth-observation systems that were built at different times by different people for diverse purposes and that use dissimilar data formats and communications techniques to operate together smoothly and from a coherent system?” 
A. “The only real answer is harmonisation”!!!

We live in exciting times for Earth Observation and the Satellite Industry in general. Just by looking at the chart of satellites launched, it is evident that the satellite industry is booming; a trend that is expected to continue its forward trajectory and keep driving an expanding space economy.

Landsat-8/9 and Sentinel-2A/B provide freely available data at an unprecedented scale and combining them to build and publish a harmonized product is a monumental challenge. And these are just four satellites.

Data interoperability with Planetscope

Planet launched their first satellite in 2013. Now, they have approximately 200 satellites in orbit (~180 Doves and 21 SkySats), capturing over 25 TB of imagery daily!

Having that large a fleet of satellites, it’s not a surprise that they have a heavy presence in workshops like the ARD Satellite Data Interoperability Workshop and JACIE. 
Many of their talks over the last few years have been focused on their CalVal processes and the challenges they face, as they have put huge efforts into improving their delivered imagery.

“Interoperability IS usability”

Planet’s fleet of Dove satellites offers global daily coverage and people have been trying to make the most out of this catalogue. 
Many studies have used Planetscope data to fill the gaps between Landsat-8/Sentinel-2 observations and create dense time series, compare crop yield models that utilise different source data (Landsat-8/Sentinel-2/PlaneScope) or even sharpen Sentinel-2 images using 3m Planetscope data¹⁰.

However, most studies have either used a limited number of Planetscope images, or have blindly used Planet’s Level 3B analytic product. Studies such as [11], that used a combination of Planetscope/Sentinel-2/Landsat-8 imagery for crop yield estimation, had even identified the differences in surface reflectance and indices values among the different sensors, without going a step further to provide an explanation for this inconsistency. In [12], the authors studied crop yield using Planetscope data, but it seems incomplete. They developed a method based on Planetscope data but the comparison of their Random Forest model using Sentinel-2 data is “unfair”. Moreover, Random Forest models aren’t the best method to compare different data sources and to decide whether “Planetscope data are better for this task than Sentinel-2”.

It is evident that in order to derive higher level products, develop applications and make fair comparisons of among methodologies, researchers need an interoperable and harmonized dataset; a harmonized product of Planetscope data based on a “gold reference” (Sentinel-2 / Landsat-8/9).

Towards this goal, Planet has been focused on their Planet Fusion ARD product, which they have presented in recent workshops.

When operating that large a fleet of satellites there are two sets of challenges:

1. Cross-calibrate to a reference sensor (e.g. Sentinel-2)

2. Ensure intra-fleet consistency ; hundreds of cubesats with potentially different relative spectral responses

During JACIE 2023, Alan Collison et al., presented Planet’s Calibration Methodology¹³.

Planet’s calibrations are based on a collection of near simultaneous crossovers between each SuperDove and Sentinel-2 (as a reference satellite). See the Figure below.

And in order to validate calibration corrections and improve consistency within the SuperDove fleet, they analyze SuperDove crossovers.

The Planet Fusion ARD Product

“The Planet Fusion processing translates original PS Top Of Atmosphere (TOA) reflectance inputs into Surface Reflectances (SR) consistent with Landsat 8 and Sentinel-2 clear-sky observations.”¹⁴

Over the last few years, folk at Planet have been trying to make use of the latest studies and frameworks (CESTEM, FORCE⁶) and combine several input data sources (MODIS/VIIRS/S2/L8) to derive a harmonized product that is consistent against other Surface Reflectance products (e.g., NASA HLS, ESA sen2cor, sen2like, USGS LaSrc).

Planet Fusion performs these top level steps:

• Radiometric harmonisation

• Geometric harmonisation

• Cloud/Shadow masking (QA product)

Check out the processing chain of the Fusion product in the slides below.

We can grasp the scale of this challenge in the following slide as a result of the variability and uncertainty in the tasks involved¹⁵.

Even subtasks in the processing chain, like estimating Aerosol Optical Depth (AOD) and robustly detecting clouds and cloud shadows are non-trivial. Errors made in the processing chain propagate to the end product affecting its quality.

I have dedicated a large part of this section to discussing Planetscope data and the efforts to achieve interoperability with other surface reflectance products such as the HLS.

I felt this was essential to further highlight the significance of harmonized data in Earth observation. Planetscope’s extensive fleet of satellites, capturing global daily coverage and producing vast amounts of imagery, has tremendous potential for driving research and applications in various fields, including agriculture, environmental monitoring, and land cover analysis.
By exploring the challenges faced by Planet in achieving interoperability with other surface reflectance products, we gain insights into the complexities of harmonizing data from diverse sensors and sources. This in-depth examination sheds light on the crucial role of harmonisation in making data from multiple sensors compatible and standardised for further analysis and application development.

Section 2

Looking Ahead: The Water’s Edge of the USGS National Land Imaging Program

Building the Global Commons of Earth Observation Data

Last year, the Landsat Advisory Group conducted a study exploring earth observation products, including data, algorithms, and workflows.³ Their findings emphasized the significance of ARD products in achieving a harmonized and standardised collection of earth information. In its findings, they argue that the explosion in the accessibility of earth observation data combined with the rise of ARD Harmonisation products, open the possibility to form, what they refer to as, the “Global Land Imaging ARD commons”.

The aim of this concept is to combine data coming from disaggregated sensors (thermal, multispectral, hyperspectral or RADAR sensors) that are flown on government or commercial satellites, into a coherent, standardised and consistent record of Earth’s change.

As described in the paper, this platform of “commons” is not just refering to data commons but also a commons of standards that facilitate multi-sensor harmonisation for end-users.

To achieve harmonisation, it is essential to establish standards that guide data providers in cleaning and aligning their data to the ‘gold’ standard. Ideally, there should also be standards set at lower quality levels to accommodate sensors of differing quality than Landsat and Sentinel-2. This ensures that diverse sensors can be incorporated into the global commons with clear communication of their quality levels.
An important aspect of these standards is their sensor-agnostic nature, working at a higher level to allow any new sensor meeting appropriate calibration targets to seamlessly integrate into the global commons of analysis-ready earth observation data. This inclusivity ensures that the initiative remains dynamic, accommodating a wide range of sensors from different sources.

The study identified NASA’s calibration resources (such as the Goddard AERONET) as a key component of making Landsat the “gold” standard and envisions a “global sensor calibration commons” that enables interoperable data collections.

Interoperability and harmonisation in this ARD platform of “commons”

Before reading this paper I had overlooked the nuanced differences between interoperability and harmonisation.

As the paper mentions, two ARD products can be interoperable and not harmonized! 
For example, Planet’s SuperDoves capture, among other spectral bands, the “Green I” and “Yellow” bands, both of which have no Sentinel-2 equivalents. Instead of excluding a piece of data from the Global Land Imaging record, the approach of this proposed platform of commons is to take advantage of the diverse spectral characteristics of each sensor, as “any measurement in the passive optical EM spectrum is relevant to earth’s changes over time”. 
According to the study, this spectral and spatial diversity is what increases the overall value of the product, as each piece of information contributes to the same mission:

Build a coherent, consistent record of Earth’s change.

“Similar diversities are present when considering Landsat/Sentinel-2. Landsat OLI and Sentinel MSI sensors have unique spectral bands that add to the diversity of any product derived from the harmonized and/or fused sources that include Landsat.”

Moreover, when considering this dynamic between interoperability & harmonisation, the study proposes the scheme HLS + X, “where X is virtually any other sensor that shares the same mission, easily incorporating diverse spectral characteristics into a harmonized output.”

Creating and implementing this platform of commons presents significant challenges. However, the HLS+X interoperable data cube holds immense potential, providing users with unparalleled insights into Earth’s changes like never before.

Further Discussion on Interoperability, Harmonisation, Fusion and ARD for ML

As I’ve already mentioned, while reading the material related to the topic and after completing a draft of Section 2 I started thinking about Interoperability, Harmonisation and Fusion and the subtle differences among them. In some cases these terms are used quite loosely in literature, leaving room for interpretation. 
Therefore, clarifying these concepts is crucial in enhancing our comprehension of the distinctions between various data products and the essential processes involved.

Interoperability, Harmonisation, Fusion

Interoperability
Given the information presented in the sources of this post, an interoperable product relies on geometrical consistency; that is, common geometric corrections are applied to all data sources that comprise the overall data product. 
The main characteristic of an interoperable product is the ability to examine any pixel in time without experiencing missalignment issues.

Harmonisation
As mentioned earlier, a product can be interoperable without being harmonized (if geometrically consistent, a combination of Sentinel-1 and Sentinel-2 images can form an interoperable data cube!).
So, to create a harmonized product, interoperability is a prerequisite.

Here, the main characteristic of a harmonized product (comprised of at least two different sensors) is being able to treat the time series data as if they originated from a single sensor.

To achieve this harmonisation consistency, there are several steps needed, such as making spectral band adjustments and applying atmospheric corrections using common radiative transfer models and methods.

For obvious reasons, harmonisation assumes a certain level of compatibility among the different sensors combined. For example, a Sentinel-1 (SAR) image can not be harmonized to a Sentinel-2 (Optical) image. And even two images coming from two different optical sensors require compatibility (at least spectrally). For example, the yellow band captured by Planet’s Doves has no Sentinel-2 spectral equivalent to use as reference and therefore cannot be harmonized. (Some have tried to derive synthetic bands when not provided “natively” by the sensor; in [16] the authors derived synthetic Planetscope red-edge and SWIR bands using linear regression of the Planetscope visible and NIR bands with the Sentinel-2 red-edge and SWIR bands).

Fusion
In this post, there are two references to “fused/fusion” products; Sen2Like Fused Surface Reflectance Products (Level 2F), where Landsat-8/9 images are upsampled to the 10/20m native resolution of Sentinel-2 and Planet Fusion.
However, in literature, the word fusion often relates to a method that combines two different data sources; this might be a method that combines different modalities, such as SAR and optical, or even input images with different spatial resolution and/or spectral bands.

Harmonized data in the context of Machine Learning
Harmonized data is of paramount importance when it comes to training ML models. 
Inconsistent data from different sensors with varying spectral characteristics and spatial resolutions can introduce noise and confounding factors that hinder the model’s ability to identify reliable patterns and make accurate predictions. On the other hand, harmonized data ensures that the input features are standardised and aligned across different sensors, making it easier for ML models to grasp underlying patterns, detect changes over time, and achieve better generalisation.

“ARD is a function of context”, and for the ML folk the context is ML-related.
Again, individual use-cases for different ML-based tasks might have different requirements. For example, for the task of cloud detection, users would need images with… clouds! And for agricultural application (crop yield estimation), the data cube would be comprised of cloud-free images.

This highlights the significance of anticipating diverse use-cases and offering clear and concise “usage instructions” tailored to various user types. Building upon the earlier examples of cloud detection and crop yield estimation, it is expected that both tasks would utilise training data from a harmonized ARD product. However, the inclusion of appropriate metadata would enable users estimating crop yield to effectively remove cloudy images from the original data, ensuring the analysis is based on cloud-free and reliable imagery.

Conclusions

Section 1 — Defining ARD and Its Importance

In Section 1, we have delved into the various definitions of “traditional” Analysis Ready Data (ARD). ARD is crucial for achieving interoperability, enabling users to analyse data from different sensors. 
The release of Landsat ARD and the subsequent HLS laid the foundations of ARD in Earth Observation.
Now, the efforts have been focused into integrating additional compatible optical mission sensors to the “gold” standards of Sentinel-2 and Landsat-8/9. Sen2Like and Planet Fusion have been stepping towards that direction, facing several challenges along the way.

Section 2 — Global Land Imaging ARD commons

In Section 2, we have highlighted some of the topics of the Landsat Advisory Group’s study that gave us a glimpse into its vision of ARD and what implementing this “Global Land Imaging ARD commons” entails.

It’s a set of commons: 
a commons of EO data, 
a commons of standards and 
a commons of global sensor calibration

where “any new sensor can add to the overall measurement of what is happening on earth”.

Having disaggregated sensors contribute to this platform of commons and the introduction of HLS+X expands the concept of ARD and makes it diverge from our original understanding.

Instead of having a coherent, consistent, harmonized datacube “ready-to-be-explored”, in this new HLS+X datacube we have every available piece of data for an area in time.

This, to me, is a major shift and it makes a few related concepts such as “interoperability, harmonisation, fusion” more entangled.

Harmonizing planet data

The pursuit of harmonisation between Planetscope data and other surface reflectance products aligns with the vision of creating a harmonized Global Land Imaging Analysis Ready Data (ARD) commons. Such an initiative holds the promise of fostering a dynamic and collaborative data ecosystem, where harmonized data sets can be seamlessly integrated and utilised by researchers and stakeholders across the globe. Ultimately, by understanding the efforts made towards interoperability and harmonisation, we gain a clearer understanding of how standardised data can unlock the full potential of Earth Observation and lead to groundbreaking advancements in understanding our planet’s changes and challenges.

Going Beyond Analysis Ready Data

Both the “traditional” and “expanded” versions of ARD serve as crucial foundations for deriving higher-level products in Earth Observation. 
The concept of ARD has proven its significance in achieving data interoperability, consistent time-series analysis, and time efficiency for various user applications. 
However, going beyond ARD opens up new possibilities for the Earth Observation community. By integrating additional compatible optical mission sensors, as demonstrated by Sen2Like and Planet Fusion, the Global Land Imaging ARD commons envisions a dynamic and collaborative data ecosystem. This platform of “commons” not only includes data but also encompasses standards and global sensor calibration, enabling diverse sensors to contribute to a coherent and consistent record of Earth’s changes.

Looking ahead, embracing the idea of a data cube that accommodates various sensors and spectral characteristics will undoubtedly push the boundaries of Earth Observation and lead to groundbreaking advancements in understanding our planet’s complexities and addressing pressing global challenges.

Afterthoughts

Redefining ARD

Throughout this exploration of ARD in Section 1 and the expanded version in Section 2, it becomes evident that the definitions and scope of ARD have evolved significantly.
As we encounter different interpretations and applications of ARD, a crucial question arises:

Should both the “traditional” and “expanded” versions of ARD be labeled under the same term?

The divergence in definitions and objectives calls for a re-evaluation of the term “ARD” itself. While the original concept served as a foundation for ready-to-use data, the expanded version embraces a Global Land Imaging ARD commons with a focus on data interoperability and inclusivity of diverse sensors. Perhaps it is time to consider redefining ARD to encompass its evolving nature and to differentiate between these distinct but complementary concepts.

A Recipe-Based Approach for ARD?

While researching the topic and writing this article, I couldn’t help but consider whether ARD could be effectively implemented through a recipe-based approach, utilising configuration files with distinct models and parameters for various stages of data processing? This recipe-concept envisions a modular workflow where raw data undergoes a series of internal products and transformations, eventually leading to the creation of the final ARD product; custom-made to the specific use-case with full traceability. Similar to following a recipe while cooking, each step would be guided by predefined configurations, making use of open-source methods and tools, akin to some of the configurable processes employed by FORCE and Sen2Like. 
By adopting such a recipe-based workflow, users could experiment with different combinations of algorithms and parameters, providing greater flexibility and customization in generating ARD products that align with specific use-cases. 
Additionally, embracing algorithm sharing and social coding principles could foster collaborative development within the Earth Observation community, encouraging experts to openly contribute to the improvement of algorithms and data processing techniques. 
However, balancing the intellectual property (IP) value of proprietary algorithms with the openness of open-source code would require careful consideration to ensure transparency, reproducibility, and accessibility of the ARD recipes and associated tools.

Feedback and corrections

While I’ve tried to accurately represent the insights and ideas from the presentations and sources discussed in this article, the complex nature of the subject may have led to unintended misinterpretations or omissions. 
To both the contributors of the materials referenced and readers, I invite you to share any corrections, clarifications, or additional context that could enhance the accuracy and depth of this article.

Acknowledgments

Many thanks to Gary Crowley and Katerina Bakousi for their invaluable feedback, insightful comments and engaging discussions that significantly contributed to the refinement of this article.

Sources

[1] Sentinel-1 ARD Normalised Radar Backscatter (NRB) Product [link]

[2] Dwyer, John L., et al. “Analysis ready data: enabling analysis of the Landsat archive.” Remote Sensing 10.9 (2018): 1363.

[3] Looking Ahead: The Water’s Edge of the USGS National Land Imaging Program — Building the Global Commons of Earth Observation Data [link]

[4] Deconstructing Analysis-Ready Data [link]

[5] Claverie, Martin, et al. “The Harmonized Landsat and Sentinel-2 surface reflectance data set.” Remote sensing of environment 219 (2018): 145–161.

[6] Framework for Operational Radiometric Correction for Environmental monitoring (FORCE) [link]

[7] https://github.com/senbox-org/sen2like

[8] Saunier, Sébastien, et al. “Sen2Like: Paving the Way towards Harmonization and Fusion of Optical Data.” Remote Sensing 14.16 (2022): 3855.

[9] Planet Fusion Monitoring — Rasmus Houborg, Planet [link]

[10] Li, Zhongbin, et al. “Sharpening the Sentinel-2 10 and 20 m Bands to Planetscope-0 3 m Resolution.” Remote Sensing 12.15 (2020): 2406.

[11] Tunca, Emre, Eyüp Selim Köksal, and Sakine Çetin Taner. “Silage maize yield estimation by using planetscope, sentinel-2A and landsat 8 OLI satellite images.” Smart Agricultural Technology 4 (2023): 100165.

[12] Nieto, Luciana, et al. “Impact of High-Cadence Earth Observation in Maize Crop Phenology Classification.” Remote Sensing 14.3 (2022): 469.

[13] Challenges for the Radiometric Interoperability of a Mixed Fleet of Medium Res, High Res and Hyperspectral Satellites — Alan Collison, Planet [link]

[14] Planet Fusion Monitoring Technical Specification, Version 1.0.0, March 2022 [link]

[15] Per-Scene Harmonization & Normalisation of Planetscope Data — Joe Kington, Planet [link]

[16] Li, Zhongbin, et al. “Sharpening the Sentinel-2 10 and 20 m Bands to Planetscope-0 3 m Resolution.” Remote Sensing 12.15 (2020): 2406.