Deep learning still has to earn its keep against boring old gradient boosting - at least for classifying city land from aerial photos.
Researchers benchmarked classical machine learning models (Logistic Regression, SVM, Random Forest, XGBoost, CatBoost) against newer tabular deep learning models (TabNet, FT-Transformer, TabTransformer, TabSeq, and 1D CNNs) on the UCI Urban Land Cover dataset. The data sorts pixels from high-resolution aerial imagery into nine classes, like roads, trees, grass, and water. It is a messy dataset by design, with high dimensionality, mismatched feature types, and uneven class sizes. The team used weighted cross-entropy loss to help the deep learning models cope with that imbalance, then compared everyone on accuracy, precision, recall, F1, and AUC-ROC.
The finding: tree ensembles are still solid default choices, but they are not automatically the best choice. Tabular deep learning models matched or beat them when the data's patterns were non-linear and the imbalance handling actually worked.
That is a narrower claim than the deep learning hype cycle usually makes, and it matches what practitioners already suspected - on structured, spreadsheet-like data, boosted trees remain stubbornly hard to beat without real engineering effort on the neural net side. Urban planners and remote sensing teams picking a model for land cover maps should read this as permission to keep using XGBoost by default, not a mandate to switch.