Early Classification Priced in Downlink Bytes

Earth observation satellites collect far more imagery than they can send to the ground. One estimate quoted in Earth+ is that only about 2% of captured data is ever downlinked. Crop-type mapping needs a full season of repeated images of each field, so a map built on the ground from whatever arrives is short of data, and it also arrives late. This proposal puts a small recurrent network on the satellite. A recurrent network is a sequence model that carries a compact memory forward from one image to the next. For each field, at each revisit, the network decides one of two things. It can "retire" the field: send a label of a few bytes and never spend downlink on that field's pixels again. Or it can keep the field "open": send this observation down so a large ground-side model can classify it from the full record.

The core idea is to train this retire-or-continue decision with a loss that weights each time step by what continuing actually costs, which is downlink bytes under a per-pass capacity budget. Existing early crop classifiers weight time steps by the fraction of the season elapsed. Under this proposal, easy fields drop out of the downlink queue early, and the scarce capacity goes to the ambiguous and rare-class fields that the ground model can resolve and the on-board model cannot.

Core Claim

Early classification of satellite time series already exists. Early Classification for Agricultural Monitoring from Satellite Time Series attached a stopping head to a Long Short-Term Memory (LSTM) network and trained it with an earliness reward scaled by $(1 - t/T)$. It reached 81% accuracy using 40% of the season, against 83% with the full year. ELECTS, in End-to-end learned early classification of time series for in-season crop type mapping, made this end to end and reached 86% overall accuracy on BavarianCrops while using only 32% of the sequence on BreizhCrops. Both papers motivate and supply the method here. Both also assume that every observation is already stored centrally, that waiting has a cost only in calendar time, and that the early classifier is the final answer.

This proposal changes all three assumptions. Waiting costs bytes, those bytes are shared across all fields in a pass, and a field that is never retired is not dropped: it is handed to a stronger ground model. Let $p_t$ be the on-board stopping probability at revisit $t$, so the probability of retiring at $t$ is $\pi_t = p_t \prod_{s<t}(1-p_s)$. Let $\hat y_t$ be the on-board prediction, $\hat y^{g}$ the ground model's prediction from the downlinked record, $b_t$ the bytes needed to send revisit $t$, and $\ell$ cross-entropy. The per-field loss is

$$ \mathcal{L} = \sum_{t=1}^{T} \pi_t\, \ell(y,\hat y_t) \;+\; \Big(1-\sum_{t}\pi_t\Big)\,\ell(y,\hat y^{g}) \;+\; \lambda \sum_{t=1}^{T} \Big(1-\sum_{s\le t}\pi_s\Big)\, b_t . $$

The stopping distribution $\pi_t$ is the temporal loss weight. Continuing is penalized by the bytes it consumes, not by the elapsed fraction $t/T$. ELECTS is roughly the special case of constant $b_t$ with no deferral term. The byte price $\lambda$ is not a hand-tuned trade-off. It is a dual variable, adjusted during training so that the expected bytes per pass meet a stated capacity. This addresses the hyperparameter-sensitivity problem that Open challenges for Machine Learning based Early Decision-Making research raises for earliness weights. Cloudy revisits carry the same byte cost but almost no information. The model is therefore pushed to learn when an observation is worth paying for, rather than treating every date alike.

Research Program

The on-board model is a small gated recurrent unit (GRU) network run per field on per-band field statistics. Its state size is fixed however long the season runs, which is the practical reason to prefer recurrence over attention on a satellite. Multi-Temporal Land Cover Classification with Sequential Recurrent Encoders supplies two pieces of evidence for this choice. Recurrent gates learned to suppress cloudy dates on raw top-of-atmosphere data, and their overall accuracy was about 89.5%. EdgeDRNN shows, but only on speech, that recurrent updates can be made cheap by skipping small state changes. We treat that as an option to measure, not a claim: we will report operation counts with and without delta-thresholded updates, and we will not report hardware latency or power figures we have not measured.

The ground model is the strongest full-sequence classifier available on each dataset. It is trained only on data it would actually receive, meaning complete sequences for fields that stay open. It is also trained jointly with the on-board policy, so that it specializes on the fields the policy defers. Contact schedules and per-pass capacities are simulated from orbital revisit patterns, with regimes ranging from one ground station to many. Capacity is set as a fraction of the full data volume, swept from roughly 1% to 20%. The 2% figure from Earth+ is used as a headline operating point. The main analysis is the accuracy-versus-bytes frontier, broken down by class frequency. We expect any gain to come from the long tail, where BREIZHCROPS documents severe imbalance and small models are weakest.

Baselines

The do-nothing reference is centralized classification from a random, uniformly subsampled set of downlinked observations at the same byte budget, fed to the same ground model. This is what a capacity-limited ground pipeline gets today. Beating it at matched bytes is the minimum bar.

The strongest existing method is ELECTS run entirely on board. It sends only labels, so it uses near-zero raw downlink. Its reported 86% overall accuracy on BavarianCrops and its 32% sequence usage on BreizhCrops are the published reference points. We retrain it with the authors' earliness loss on identical splits, with the same GRU capacity and the same tuning budget as our on-board model. This is the most dangerous competitor. If the small on-board model already matches the ground model, then sending labels only is optimal, and the ground-deferral branch adds nothing. Our method beats it only if, at some byte budget, it improves macro-averaged accuracy or macro-mIoU over ELECTS, with no loss in overall accuracy.

A second strong competitor is a fixed-rule cascade: ELECTS on board, with every field not retired by a confidence threshold simply deferred. The threshold is tuned on validation data to meet the same budget. This isolates whether training with byte-priced weights matters or whether any deferral rule does as well.

Two ablations isolate the key ingredients. The first replaces $b_t$ with a constant, which recovers time-priced earliness plus deferral. The second removes joint training of the ground model. For the ground model, the published reference points are the Transformer's roughly 0.80 overall accuracy on BreizhCrops and TSViT's 48.1% macro-mIoU on SwissCrop25. No reported result was found on Lacuna for any method evaluated by accuracy at a fixed downlink byte budget, so all numbers on that frontier will be new.

Why not just…?

Why not just compress everything and classify on the ground? Learned feature compression such as FOOL reports 51% lower bitrate than earlier bottleneck methods with under 1% loss in detection mAP. That is a constant-factor saving on every image. Retiring a field removes all of its later revisits, which is a saving that grows over the season. The two approaches combine rather than compete: we will add a compressed-feature variant of the open-field downlink and report whether retirement adds savings beyond compression alone.

Why not just run ELECTS on board and send labels? This is a real threat. On balanced, frequent classes it may win outright. The experiment settles it by class stratum. If the ground model's gain on rare classes at full data does not exceed the on-board model's by a clear margin, deferral cannot help, and we will report that as the finding.

Why not train with uniform loss and stop when softmax confidence is high? The earliness-reward papers above report that learned stopping heads beat post-hoc rules. However, neither tested against a budget-matched confidence cascade with a ground fallback. The fixed-rule cascade baseline tests exactly this. If it matches our method, the byte-weighted loss is unnecessary.

Why not a transformer on board? TSViT leads SwissCrop25, but that benchmark notes its advantage appears only late in the season. Full-sequence attention also needs the whole history in memory. We use the transformer on the ground, where memory is not the constraint, and compare it against the on-board GRU at full data to measure what deferral can buy.

Proposed Experiments

The first experiment uses BreizhCrops, BavarianCrops, Ghana, and South Sudan, with ELECTS's splits. We trace accuracy, macro accuracy, and median label-delivery date as functions of downlink budget for all baselines.

The decisive check is whether the byte-priced policy's frontier lies above both ELECTS-on-board and the budget-matched confidence cascade. It should do so specifically at budgets near 2%, and the gain should concentrate in classes below median frequency. A second check replays the simulation with cloud-heavy revisits. The policy should learn to skip paying for them, which we verify by comparing the per-date retain rate against cloud fraction.

The stress test uses SwissCrop25's leave-one-year-out split, aggregated to parcels. The question is whether a byte price learned in normal years causes premature retirement in shifted years. We compare calendar-day time indexing with growing-degree-day indexing, which that benchmark found mitigates cross-year shifts, and ask whether the second restores the frontier.

The idea fails cleanly if ELECTS-on-board matches our frontier on macro metrics at every budget. That outcome would show that on-board early classification alone solves in-season mapping under downlink scarcity.