A Reliability-Aware Edge–Ground Cascade for Cloud-Occluded Satellite Scene Classification

Selective offloading—answering confident inputs on a cheap edge model and deferring only unreliable ones to a stronger model—is a standard tool for cost-sensitive inference, and its accuracy–cost trade-off hinges on a single design choice: which per-input reliability signal to threshold. Softmax confidence is the default and a famously hard-to-beat baseline, and a recurring question is whether richer uncertainty- or shift-sensitive signals do better once inputs are corrupted. We study this question in a concrete, operationally important instance: on-board satellite scene classification under cloud occlusion, where clouds are a real, physically-grounded distribution shift and downlink bandwidth is genuinely scarce, so unreliable tiles are offloaded to a stronger ground model. We benchmark seven signals against confidence as offload triggers—attention-drift (Grad-CAM dispersion), predictive entropy, margin, max-logit, an energy out-of-distribution score, temperature-scaled confidence, and cloud brightness. Across two remote-sensing benchmarks (RESISC45, EuroSAT), four backbones (CNNs and a mobile transformer), and three occlusion models including real 38-Cloud composites—a cross-sensor approximation, since these are Landsat-8 clouds blended onto aerial/Sentinel-2 tiles—we find that no signal meaningfully beats confidence as a standalone trigger: margin, max-logit and entropy are statistically tied with it, while attention-drift and cloud-brightness are far worse. Fusing attention-drift with confidence helps only narrowly—at the heavy-cloud cascade operating point on the harder RESISC45 task under real cloud composites, where it cuts offload bandwidth by up to 10 points at fixed accuracy—and is neutral elsewhere; we show its apparent AUROC-level benefit is largely an under-training artifact that mostly vanishes at convergence. We further show occlusion degrades both accuracy and calibration, and that a confidence-triggered cascade saves up to 37 percentage points of bandwidth versus naive offloading. Our contributions are a controlled, honestly-reported negative for richer selective-offloading signals under a real distribution shift, a reproducible reliability-signal benchmark, and a deployable confidence-triggered edge–ground cascade, together with a practical warning that FLOPs mispredict on-board latency.