Accurate land cover classification in fragmented agricultural landscapes remains a persistent challenge, demanding robust algorithms capable of exploiting both spectral and temporal complexity. This study presents a rigorous, controlled comparison of traditional machine learning and deep learning frameworks for multi-temporal land cover classification using Sentinel-2 imagery over the Bafra Plain, northern Türkiye, a highly heterogeneous agro-ecosystem with sub-hectare parcels and pronounced phenological variability. A temporally enriched dataset of 78 spectral-temporal variables, derived from ten Sentinel-2 bands and three vegetation indices (NDVI, NDWI, NDRE) across six acquisition dates, was constructed to capture intra-annual phenological dynamics. To ensure spatial independence and eliminate data leakage, a parcel-based stratified sampling scheme produced strictly non-overlapping training and testing partitions of approximately 2.94 million pixels each. Support Vector Machines (SVM), Random Forest (RF), a baseline Convolutional Neural Network (CNN), and U-Net were benchmarked under identical conditions using Overall Accuracy (OA), class-wise F1-score, Cohen's Kappa coefficient, and mean Intersection over Union (mIoU). U-Net achieved superior performance across all metrics (OA: 96.87%; Kappa: 0.957; mIoU: 89.42%), surpassing SVM by up to 5.09 percentage points and exceeding the widely adopted 85% mIoU excellence threshold, which only the CNN (86.84%) and U-Net reached, whereas SVM (81.15%) and RF (84.62%) fell below. Gains were most pronounced for spectrally ambiguous minority classes, including Artificial Surfaces (+15.25 pp F1-score vs. SVM), Tree Crops (+13.15 pp), and Grassland (+16.21 pp), underscoring the critical role of spatial context modeling. A systematic multi-temporal ablation study confirmed that temporal integration is indispensable, yielding accuracy improvements of 4.79–7.66 percentage points across all model families relative to single-date baselines. Although deep learning entails greater training overhead, its near real-time inference and robustness to severe class imbalance render it operationally viable for large-scale monitoring, providing evidence-based guidance on the accuracy–efficiency trade-off for algorithm selection in precision agriculture and environmental monitoring.
ALTAY et al. (Fri,) studied this question.