A self-supervised MoCo v2 (He et al. 2020; Chen et al. 2020) encoder pretrained on 50000 16x16 raster patches sampled from a core Cerrado AoI (longitude -53 to -43, latitude -23 to -10), covering the Brazilian states of Goias, Tocantins, Mato Grosso, Bahia and Minas Gerais. The input stack is aligned to a 0.01-degree (~1 km) grid and combines three public keyless sources: - SoilGrids 250m, 0-5 cm mean: SOC, clay, sand, pH(H2O), bulk density (5 layers) - WorldClim 2.1 (Brazil country pack): 12 monthly precipitation + 12 monthly mean temperature (24 layers) - SRTM 30 arc-second: elevation + slope (2 layers) The encoder is a 5-block convolutional backbone producing a 64-dimensional feature vector followed by a 2-layer MLP projection head (feature_dim = 64, proj_dim = 32). Training uses a queue of 4096 negatives, InfoNCE temperature 0.07, momentum 0.999, Adam learning rate 3e-4, batch size 64, for 20000 optimisation steps on an Apple Silicon M1 Max via torch::backend_mps. Artefact files: encoder_q.pt (state_dict, SHA-256 44ace7f78c658b6028f1cf5ccfa624023295e5576f681d0135db64726c6738e8), metadata.json (full training configuration), loss_history.rds (per-step InfoNCE loss), encoder_q.pt.sha256 (digest sidecar). Consumed by the edaphos R package (>= 1.2.0) via foundation_weights_load('edaphos-cerrado-moco-v1').