On-orbit model training for satellite imagery with label proportions

This work targets training machine learning models on orbiting platforms, generally constrained by limited on-board hardware capabilities and uplink bandwidths. We aim at enabling orbiting spacecrafts to (1) continuously train lightweight models as they acquire imagery and (2) receive new labels while on orbit to update their training tasks. For this, we consider chip-level regression tasks (i.e. predicting the vegetation percentage of a 20 km² image chip) with even coarser label proportions, such as municipality-level statistics regularly available in many regions and applications. This is a learning from label proportions (LLP) problem. LLP applied to Earth observation (EO) data is an emerging field, where comparative studies remain challenging due to the lack of standardised datasets. We show how very simple methods (with ~5K parameters) outperform more complex ones, better inferring features from coarse geometries to the image chips contained within. Then, we publish a set of benchmarking datasets enabling comparative LLP applied to EO, providing both fine-grained and aggregate labels according to existing administrative divisions. Finally, we show how this approach fits on-orbit training scenarios by vastly reducing the number of computations and the size of the label sets uplinked. Source code is available at https://github.com/rramosp/llpeo.