Sensing Before Modeling: The Counterintuitive Logic of Earth Observation System Design

In Earth Observation (EO) projects, the typical starting point is picking a model — someone has a segmentation network that runs on satellite imagery, and the entire system architecture grows around it. But three agricultural and Earth observation architectures from CodeNinja Atoms point to the opposite conclusion: what determines "what to observe" is never model accuracy, but legal permissions, revisit cycles, and data ownership. More counterintuitively, two of the three systems have empty model registries — and that's entirely by design.

The three systems are deployed in Pakistan, Saudi Arabia, and the United States, all covering the same industry (agriculture and Earth observation), yet each presents a strikingly different face when it comes to the most fundamental sensing decisions. What they share: sensing decisions come first, driven by constraints that have nothing to do with accuracy.

Two of Three Systems Have Empty Model Registries — That's a Conclusion, Not a Gap

Pakistan's Field Ledger is an agricultural data dashboard built on twelve objects. Its model registry records zero models, and the reason is spelled out plainly: the requirement itself only calls for aggregation, drill-down, and reporting — no models were validated in this round. Predictors based on sensing time series are explicitly flagged as "open future work," gated behind Milestone 1 evidence rather than baked into the design as assumptions.

The US Baseline is a one-time flight acquisition project capturing four-band orthoimagery and LiDAR, built on sixteen objects. Its registry is also empty, but for a more direct reason: there is no need for a model layer at all. Vegetation analysis is left to the operator's own team using the operator's own tools. The entry that follows is one most model-centric designs would never write — since there are no weights, there are no model licenses to hold or trigger, and IP relationships therefore run through a service agreement rather than a model license.

The only system that registers any models is Saudi Arabia's Fodder Watch, which lists two, built on fourteen objects: a detector, RF-DETR (Apache 2.0, fine-tuned by region), and a predictor, Chronos-2 (Apache 2.0, zero-shot, approximately 120 million parameters). Both licenses are part of the decision, not an afterthought appended to it.

An empty registry is a conclusion, not a blank. It signals that cross-system data joins are the real hard problem, and that a learning component is not needed to solve it.

What Actually Determines the Sensing Approach Has Never Been a Benchmark Score

Every design's sensing entry carries its own rationale — and not one of those rationales is a leaderboard result.

Fodder Watch needs to screen parcels for restricted cultivation. Its optical channel is Sentinel-2, 10-meter resolution with a 5-day revisit, chosen because it can distinguish crop types under center-pivot irrigation without relying on a foreign processing chain. Its radar channel is Sentinel-1 C-band SAR, and the stated reason is worth reading twice: this cloud-resistant supplement matters because in that geographic environment, the real enemy of the optical channel is dust and haze, not clouds. Landsat 8/9 (8-day combined revisit) is there for its archive stretching back decades — precisely what gives each parcel a baseline time series. Cloud cover from Sentinel-2 is recorded as "evidence" within the screening rules, not as "ground truth."

Baseline uses a manned aircraft carrying a four-band (red, green, blue, near-infrared) large-format camera — because the requirement explicitly rules out drone data. Its LiDAR is a Quality Level 1 (QL1) sensor with a minimum of 8 pulses per square meter, with a 16-pulse option separately priced and preserved as an exercisable right. What actually determines whether the deliverable qualifies as QL1 is the positioning entry: GNSS navigation with network RTK or a private base station, specified by baseline, correction source, and test evidence — because accuracy claims rest entirely on that chain.

Field Ledger installs no sensors at all. It only reads from sensing equipment the operator already has — read-only, no cameras, no new field hardware. The world is perceived "per farm," not surveilled.

Three systems in the same industry, with sensing sources ranging from four classes of satellite and aerial data down to zero. That span is itself the argument: sensing is a requirements decision.

Technical Background: Sentinel-1/2 and Landsat

Sentinel-1 and Sentinel-2 are freely and openly available satellites under the European Space Agency's Copernicus programme. Sentinel-2 carries a multispectral sensor with 10-meter resolution across visible and near-infrared bands; a two-satellite constellation achieves approximately 5-day revisit. Sentinel-1 carries a C-band Synthetic Aperture Radar (SAR), capable of penetrating clouds and dust to capture surface backscatter, with a revisit period of roughly 6–12 days. SAR's imaging principle is fundamentally different from optical sensors — it actively emits microwaves and records the echo, requiring no solar illumination, which gives it an irreplaceable advantage over optical satellites in arid regions prone to sandstorms or haze. This is the physical reason Fodder Watch lists it as a "supplementary" channel. Landsat 8/9, operated by the US Geological Survey (USGS), has an archive dating back to 1972, making it the longest continuous EO sequence available — 30-meter resolution, 16-day revisit (approximately 8 days with both satellites). Its value lies not in current precision but in providing a historical baseline no other sensor can replicate.

Licenses Are a Design Constraint, Not Administrative Paperwork

Fodder Watch contains a self-explanatory record: the XL and 2XL versions of RF-DETR, under Roboflow's platform model license, are excluded. The reason is that their "field of use" clause would compromise ownership of the weights within the Kingdom of Saudi Arabia.

This is a decision to reject a checkpoint on licensing grounds while retaining the smaller Apache 2.0 checkpoint. Had the model been selected first and the sovereignty requirement read afterward, the entire design would have needed to be rebuilt around a constraint that could have been satisfied on day one. The license here is not a signature at the end of the process — it is a hard upstream constraint.

RF-DETR and What Apache 2.0 Actually Means in Practice

RF-DETR is a real-time object detection model developed by Roboflow, built on an improved DETR (Detection Transformer) architecture with a strong balance between accuracy and inference speed. Its larger versions (XL, 2XL) are released under Roboflow's platform model license, which typically includes "field of use" or "deployment scope" restrictions requiring deployment through Roboflow's inference service or under specific conditions — meaning effective control of the weights is partly held by the platform, not the user. Apache 2.0 is a permissive open-source license that permits commercial use, modification, and redistribution without deployment platform restrictions; weights can be fully held on-premises. For scenarios requiring data and model sovereignty to remain within a specific jurisdiction, license type directly determines whether weights "belong" to the user — this has nothing to do with price. It is a legal constraint at the sovereignty level. Chronos-2 is a time-series forecasting foundation model developed by Amazon, also released under Apache 2.0, with approximately 120 million parameters and support for zero-shot time-series prediction.

In Some Scenarios, Renting Isn't a "Price Question" — It's Simply Off the Table

Fodder Watch needs a 48 GB PCIe card (L40S class), selected via a "memory footprint rule": combined weights under 0.55 GB, with headroom for batch processing and caching. Owning this node over three years costs approximately $15,200 (range $14,600–$15,800), factoring in 8–12% annual support fees and electricity calculated at PUE 1.6 against Saudi industrial power rates. Renting the equivalent card on-demand in the nearest AWS region runs approximately $60,000 over three years; even the deepest three-year EC2 Instance Savings Plan (full upfront) comes to $22,600.

Owning wins on cost here. But the line in the appendix that truly matters is the qualifier attached to every rental price: outside the Kingdom. For a service whose core purpose is keeping detection results sovereign, the cheapest cloud price is not a competitive option — it is simply a measurement of something the requirements already excluded.

PUE and QL1: What They Actually Mean

PUE (Power Usage Effectiveness) measures data center energy efficiency, defined as total facility power draw divided by IT equipment power draw. A PUE of 1.6 means that for every 1 watt spent on computing, an additional 0.6 watts is consumed by cooling, lighting, and other infrastructure — a moderately realistic deployment figure (hyperscale cloud providers can reach 1.1–1.2; traditional data centers commonly see 1.5–2.0). When converting to electricity costs, PUE directly affects the actual consumption calculation, making it a necessary parameter in any cost comparison.

QL1 (Quality Level 1) is the highest LiDAR point cloud accuracy tier under the USGS 3DEP standard and several industry specifications, requiring a minimum point density of 8 pulses per square meter (pls/m²) and vertical accuracy on the order of 10 centimeters. Point density directly affects the ability to resolve low-canopy vegetation, bare-ground microtopography, and similar fine detail. In the Baseline project, this is the core technical anchor for LiDAR procurement negotiations — not an optional upgrade.

The Decision Sequence These Designs Actually Follow

Reading the three systems side by side, a consistent sequence emerges:

1. What do law and the operator permit you to sense?
2. What revisit cycles and resolutions does that sensing deliver?
3. What objects must those readings become for the question to be answered?
4. Only then: does a learning model need to exist — and if so, under what license, running on whose hardware?

The model is the fourth decision. And in two out of three cases, it is not a decision at all.

This "sense first, model later" logic is a useful prompt for any team building EO systems from the model outward: accuracy benchmarks matter, but legal permissions, revisit cycles, and data ownership — the non-technical constraints — are often the things that need to be locked down on day one of design.

All three designs, including their object models (JSON), model registries, and sourced cost appendices, are openly available for reuse under CC BY 4.0.