edaphos-cerrado-moco-v2 — a 200k-step MoCo v2 foundation-model encoder for the Brazilian Cerrado soil covariate stack

edaphos-cerrado-moco-v2 is a self-supervised contrastive encoderfor the Brazilian Cerrado soil covariate stack, trained with theMoCo v2 algorithm (He et al. 2020) for 200,000 InfoNCE steps —ten times the v1 (10.5281/zenodo.19701276) optimisation budget. Architecture: 31-channel input (5 SoilGrids 250 m layers{soc, clay, sand, phh2o, bdod} + 12 WorldClim 2.1 monthlyprecipitation + 12 monthly mean temperature + 2 SRTM topographic),patch_size = 16 × 16, feature_dim = 64, proj_dim = 32, queue_size= 4096, momentum = 0.999, temperature = 0.07, lr = 3e-4. Use case: feature extractor for downstream pedometric tasks(soil organic carbon prediction, classification, instrument-variable causal inference). The v1.3.2 release of the edaphos Rpackage <https://github.com/HugoMachadoRodrigues/edaphos>benchmarks the v2 encoder against v1 to test whether a fairly-trained representation rescues the clay → SOC IV stage-1 Fstatistic above the weak-instrument threshold. Loaded in R via: library(edaphos) moco <- foundation_weights_load("edaphos-cerrado-moco-v2") Files in this deposit: encoder_q.pt — torch state dict of the query encoder encoder_q.pt.sha256 — SHA-256 digest for verification metadata.json — full training config + AoI + per-channel normalisation stats loss_history.rds — 200 000-element InfoNCE loss vector README.md — citation + load instructions Supersedes (does NOT replace): edaphos-cerrado-moco-v1(10.5281/zenodo.19701276). v1 remains the canonical referencefor any benchmark / paper citing it; v2 is a new artefact. Citation: Rodrigues, H. (2026). edaphos-cerrado-moco-v2: a 200k-step MoCo v2 foundation-model encoder for the Brazilian Cerrado soil covariate stack. Zenodo. https://doi.org/10.5281/zenodo.19831617