A Mission-Programmable Onboard Decision Layer for Earth-Observation CubeSats: Scene Triage and Vegetation-Loss Alerting from 68-Byte Classifier Weights

An Earth-observation CubeSat cannot downlink every scene, so what to send must be decided in orbit, yet existing onboard systems either fix their priority logic before launch or make it reprogrammable through a heavy encoder an operator cannot inspect. This work presents a mission-programmable triage system whose contribution is architectural: sixteen spectral–textural Sentinel-2 features feed a classifier-agnostic decision layer mapping each patch to four land-cover classes and driving two orbit-reconfigurable modes with no retraining. A 68-parameter logistic classifier, 68 bytes in int8 within a 150 kB executable, reaches Cohen’s κ = 0.838, 0.887 agreement, and 0.849 macro-F1 with the SCL labelling; as the SCL also supplies the labels, this is agreement with the scheme, cross-checked against an independent product (ESRI/IO, κ = 0.78). Mode A ranks scenes against an eight-byte target, cutting downlink by up to 77.7%. Mode B emits a few-hundred-byte alert reaching 88.6% precision against Hansen Global Forest Change at 28.8% recall, falling below chance on diffuse loss (Mekong–Laos) and alerting on 1.78% of sub-zones across six low-prevalence controls. At 27 s and 37 J of incremental energy per scene on a commercial ARM64 proxy, this interpretable, byte-scale decision layer turns the downlink into a selective, verifiable instrument.