Compact CNN-Transformers match larger models for spatially-robust soil organic carbon mapping at regional scale

Soil Organic Carbon (SOC) is a global reservoir in the carbon cycle, and accurate regional monitoring is needed for soil assessment and climate-change mitigation. Most Digital Soil Mapping (DSM) studies treat SOC as a static spatial target and report accuracy on random data splits, which spatial autocorrelation inflates. We evaluate a family of architectures, gated and ungated CNN-Transformer hybrids, plain Transformers, CNN–LSTM, Random Forest, and XGBoost, on 16,514 Bavarian topsoil samples (2007–2023) with a 43-band multi-temporal covariate stack, under 10-fold longitude-blocked spatial cross-validation. Every attention model clusters within one fold standard deviation, so a compact gated CNN-Transformer ( ≈ 87k parameters, R 2 = 0.375 ± 0.100 ) comes within one fold standard deviation of a ≈ 45.1M-parameter pure Transformer ( R 2 = 0.401 ± 0.097 ) at roughly 500 × fewer parameters. Neither the convolutional front-end nor the gated residual network yields a separable gain, and added capacity does not help. Tree ensembles trail only modestly on the full domain (Random Forest 0.342, XGBoost 0.317). This edge is modest (0.05–0.07 in fold-mean R 2 , within the fold-to-fold spread) but consistent in sign across longitude-, latitude-, and cluster-blocked folds, and it is land-use stratified: on cropland alone, a low-variance regime, gradient-boosted trees are far stronger ( R 2 ≈ 0.66 versus ≈ 0.21 ), reproducing prior cropland-only Bavarian studies. The appropriate regional product is therefore a land-use-stratified map with tree ensembles on cropland and attention models on the high-variance organic remainder, rather than a single model. The contribution is thus not a new architecture for SOC mapping but a rigorous, spatially-validated architecture comparison and the land-use-dependent inversion of model performance it reveals. The recommended compact model produces a wall-to-wall 2023 SOC map of Bavaria at 250 m resolution.