Reliability-Aware Evaluation of TabPFN for Small-Sample Crop Classification in Mountainous Agricultural Regions Using Multi-Temporal Sentinel-2-Derived Features

Crop classification in mountainous agricultural regions is challenged by fragmented fields, limited reference samples, class imbalance, and strong spatial heterogeneity. We evaluated the Tabular Prior-data Fitted Network (TabPFN) for reliability-aware small-sample crop classification using 2857 reference samples from Yuxi City, China, and 66 multi-temporal Sentinel-2-derived spectral, vegetation-index, and texture features, excluding longitude and latitude from model inputs. Under five fixed stratified 70/30 splits, TabPFN achieved an overall accuracy of 0.8944 ± 0.0127 and a Macro-F1 of 0.8798 ± 0.0173, outperforming five conventional baseline classifiers. Probability reliability analysis and selective classification further showed that confidence scores were informative of prediction risk and could support the prioritization of low-confidence samples for expert review. Calibration analysis also indicated agreement between confidence and empirical correctness. In matched five-fold validation, pooled Macro-F1 decreased from 0.8632 under stratified random folds to 0.6559 under projected-coordinate spatial folds, a 24.0% relative reduction, although TabPFN remained the highest-ranked classifier among eight tabular models. Excluding four EVI-derived predictors reduced Macro-F1 from 0.8806 to 0.8573, indicating a useful but non-exclusive dependence on these predictors. Overall, the results show that random validation can be optimistic under spatial dependence and that broader cross-region and cross-year generalization requires independent evaluation.