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

Accurate, privacy-respecting soil information is essential for site-specific nutrient management and carbon accounting, yet the cost of laboratory analyses limits many farms to relatively sparse sampling grids. We propose a collaborative framework that couples a national Sentinel-2 bare-soil composite with FL to produce high-resolution clay and soil organic carbon (SOC) maps while keeping all local data on-premise. A one-dimensional convolutional neural network was first pre-trained on a 53,570-sample Brazilian archive and then fine-tuned across 50 farms through synchronous Federated Averaging. We benchmarked this hybrid configuration against (i) a purely centralized model trained on the same archive and (ii) a fully decentralized FL model initialized at random. Across farm-level validation sets, pre-trained FL lowered median RMSE by 42% for clay and 31% for SOC relative to the centralized baseline, while increasing median RPIQ by 33% and 25%, respectively. On farms with ≤ 150 samples, the gains remained substantial, confirming that the approach remains effective when local datasets are modest compared with national archives. Error distributions differed significantly between scenarios (Friedman and Wilcoxon tests), and the pre-trained FL maps removed most spatial artefacts observed in the centralized outputs while preserving fine-scale gradients. Because only encrypted weight updates are exchanged, sensitive information never leaves the farm, satisfying GDPR/LGPD-style constraints and allowing late-joining clients to inherit an increasingly mature global model. Taken together, these results indicate that continental pre-training followed by federated fine-tuning reconciles global generality with local specificity and offers a scalable blueprint for privacy-preserving, high-resolution soil mapping in settings where sample densities are often substantially lower than in experimental setups, without compromising data sovereignty. • Continental pre-training + federated fine-tuning beats centralized baselines. • Shares only encrypted weight updates, keeping all laboratory data on-premise. • Clay RMSE ↓ 42 % and SOC RMSE ↓ 31 % across 50 heterogeneous Brazilian farms. • Supports data-scarce clients (<200 samples) without loss of accuracy.