Architecture diverse TabNet ensemble for crop type classification using harmonized multisource soil data

Abstract Accurate analysis of heterogeneous soil records can support data-driven agricultural planning, but integrating datasets from different laboratories and public repositories introduces challenges related to inconsistent units, source-dependent distributions, missing values, duplicate observations, and potential information leakage. This study proposes ADWTabNet, an architecture-diverse TabNet ensemble for multiclass crop-type classification using harmonized multi-source soil data. The integrated dataset contained 18,237 observations representing seven crop classes and a broad set of chemical, physical, microbiological, and categorical soil properties. The proposed framework combines unit standardization, duplicate control, source-distribution harmonization, separate categorical and numerical imputation, tree-structured Parzen estimator-based hyperparameter optimization, structurally diverse TabNet models, probability-level ensemble fusion, calibration assessment, and attention-based interpretation. Under the evaluated stratified holdout protocol, XGBoost achieved a test accuracy of 0.8724 and a macro F1-score of 0.8670. Baseline TabNet improved these values to 0.8863 and 0.8822, respectively, while the TPE-optimized TabNet reached an accuracy of 0.8951 and reduced log loss from 0.3325 to 0.2566. The proposed ADWTabNet ensemble obtained the best reported performance, with a test accuracy of 0.9053, macro F1-score of 0.9025, weighted F1-score of 0.9050, macro ROC–AUC of 0.9930, and normalized multiclass Brier score of 0.0193. Repeated-run evaluation confirmed a stable advantage over the optimized individual TabNet, whereas the difference between validation-weighted and equal architecture-diverse averaging was not statistically significant. Aggregated TabNet attention ranked humus, phosphorus, electrical conductivity, nitrogen, and pH among the leading predictors, and this ranking was independently supported by permutation-importance analysis. These findings demonstrate improved crop-type classification within the evaluated integrated multi-source dataset. However, external validation using previously unseen laboratories and geographic regions is required before broader transferability can be established.