"South Africa Crop Comp"

"Accurate crop classification underpins food-security monitoring across the Global South, yet operational mapping there means applying a model to regions where no labels were collected. Most studies instead validate within a single region (random or field-wise cross-validation), which measures in-distribution fit rather than spatial transfer. On a South African Sentinel-2 dataset (optical only), we benchmark twenty-three classical, and deep models twice: in-region and on a spatial holdout tile. The rankings invert. In-region the dense temporal deep nets lead (macro-F1 0.72--0.78); under transfer this cluster fractures, as pixel-level L-TAE and a Transformer encoder drop from 0.78 to 0.58 and CNN-BiLSTM falls the most (0.72 to 0.46), while a TabNet ensemble becomes the best model (0.60) and simple trees and logistic regression (0.56) prove robust. The results suggest a mechanistic explanation: models with sparse, axis-aligned feature-selection are more stable under transfer. TabNet inherits it through sparsemax masks, while dense nets discard it. A control isolates it: a Transformer with far greater attention capacity transfers no better than L-TAE, so capacity is not the lever; sparsity is. We graft that bias onto a deep network, L-TAE-S, which TabNet for the best spatial-transfer macro-F1, and its field-level variant has the smallest in-region-to-holdout gap of any model. Field-level aggregation is a second, substitutable lever, and robust models hold near-full accuracy on a quarter of the fields. The message is twofold: model selection in crop mapping is an artifact of the validation protocol, and the property deciding whether a model survives transfer is sparse, axis-aligned feature selection."