Sky Condition Classification on Embedded Systems Using TinyML

Abstract

Adaptive link selection in hybrid radiofrequency/free-space optical (RF/FSO) ground stations for Low Earth Orbit (LEO) satellites requires local atmospheric sensing that meets sub-second update requirements and is deployable on low-cost hardware. This work presents a Tiny Machine Learning (TinyML)-based five-class sky-condition classifier designed as the sensing stage of an adaptive RF/FSO ground-segment architecture. The proposed approach performs on-device inference without cloud connectivity and distinguishes Clear, Partly Cloudy, Overcast, Rain, and Noise conditions to support link-selection decisions under rapidly changing atmospheric conditions. Experimental results show that the embedded implementation exceeds 80% classification accuracy while meeting the sub-second update requirement derived from the RF/FSO sensing problem, with INT8 post-training quantization delivering a 7.41$\times$ latency reduction relative to the unquantized baseline at a cost of 2.7 percentage points (pp) in accuracy. These results indicate that microcontroller-class vision-based sensing is a feasible baseline for low-power atmospheric awareness in future adaptive LEO ground stations.