COMPARATIVE EVALUATION OF CLASSICAL MACHINE LEARNING AND DEEP LEARNING FOR CROP TYPE CLASSIFICATION FROM SENTINEL-2 NDVI TIME SERIES

Accurate crop type classification from satellite imagery underpins food security monitoring, subsidy verification, and precision agriculture. The objective of this study is to establish which classification paradigm – classical machine learning or deep learning – provides higher accuracy and better cross-domain transferability for crop type mapping from Sentinel-2 NDVI time series, focusing on the underrepresented conditions of Ukrainian agriculture. Nine classifiers – six classical (Random Forest, SVM, XGBoost, LightGBM, KNN, Gradient Boosting) and three deep-learning-based (Temporal CNN, Bidirectional LSTM, Transformer encoder) – are evaluated on the multi-country AgriSen-COG benchmark and independently tested on a Ukrainian farm dataset (Krynychanska community, Dnipro region; four crops, 2025 season, 38,867 samples). Performance is assessed through overall accuracy (OA), balanced accuracy, weighted/ macro F1, Cohen’s kappa, Matthews correlation coefficient, ROC AUC, and per-class F1. The scientific novelty lies in: (a) the first unified nine-model comparison under identical protocols on a standardized European benchmark with simultaneous cross-domain transfer evaluation on independent Ukrainian field data; (b) quantification of the domain gap between Western-European benchmark and Ukrainian agricultural conditions for both classical and deep learning methods. Results show that deep learning models outperform classical classifiers on the benchmark, with TempCNN achieving the highest OA (0.839) – 3.2 % above the best classical model (XGBoost, 0.813). Upon cross-domain transfer, accuracy degrades substantially across all models (mean ΔOA = -56.4 %), yet deep learning retains a relative advantage: DL models lose on average 52.1 % vs. 58.5 % for classical ML, and TempCNN transfers best (OA = 0.441, +34.9 % relative to the best classical transfer). Per-class analysis reveals that only maize transfers reliably (F1 ≥ 0.82), while rapeseed and sunflower require domain adaptation. Conclusions. The practical significance of the results consists in providing evidence-based guidelines for selecting classification methods in operational crop monitoring systems, demonstrating that TempCNN is the preferred model for both in-domain accuracy and cross-domain robustness, and confirming that domain adaptation is essential before deploying benchmark-trained classifiers in Ukrainian agriculture.