Federated earth-observation models for collaborative farm-scale soil mapping

Accurate, privacy-respecting soil information is essential for site-specific nu-trient management and carbon accounting, yet most farms can afford onlysparse laboratory sampling. We propose a collaborative framework that fusesa national Sentinel-2 bare-soil composite with federated learning (FL) to pro-duce high-resolution clay and soil organic carbon (SOC) maps while keepingall local data on-premise. A 1-D convolutional neural network was (i) pre-trained on a 53,570-sample Brazilian archive and (ii) fine-tuned across 50farms (≈ 157 samples per farm) through 100 synchronous FedAvg rounds.We benchmarked this Scenario C against a purely centralized model trainedon the same archive Scenario A and a fully local FL model initialized atrandom Scenario B.Across the validation sets of all farms, Scenario C lowered median RMSEby 42 % for clay and 31% for SOC relative to Scenario A, while increasingmedian RPIQ by 33 % and 25 %, respectively. On the most challengingfarms, clay errors fell more than three-fold and SOC errors were halved,eliminating the spatial artefacts and distributional mismatches observed inthe centralised outputs. Because only encrypted weight deltas are exchanged,sensitive information never left the farm, satisfying GDPR/LGPD constraintsand allowing late-joining clients to inherit an increasingly mature global model.These results demonstrate that continental pre-training followed by fed-erated fine-tuning reconciles global generality with local specificity, offeringa scalable blueprint for delivering trustworthy, high-resolution soil maps todata-scarce smallholders without compromising data sovereignty.