GeoFM_v1

We finalized the GeoFM v1 baseline for the Cerrado land-use transition pilot. The workflow consists of a self-supervised spatiotemporal patch pretraining stage (masked class prediction and next-year prediction) followed by multi-task fine-tuning with two supervised heads: (i) soy-related transition trajectory classes and (ii) regeneration-related trajectory classes. We saved a fully self-contained checkpoint including model configuration and label mappings, and additionally exported modular artifacts (encoder-only and heads-only) to facilitate transfer, explainability analyses, and future head replacement. Test-set evaluation artifacts (macro-F1, balanced accuracy, per-class reports, and confusion matrices) were generated and archived to provide a stable baseline for subsequent iterations (Explain v1, contextual conditioning, and scenario simulation toward 2030).