Background/Objectives: Accurate land cover mapping from multispectral Sentinel-2 imagery is fundamental for environmental monitoring, efficient natural resource management, and spatial planning. While deep learning has become the dominant approach for semantic segmentation, the combined impact of spectral input selection and network architecture on cross-regional robustness remains insufficiently explored. This study systematically investigates multispectral land cover segmentation in Serbia and evaluates its transferability to Western Balkan regions using a structured experimental framework. Methods: A comprehensive band-combination ablation analysis (3–10 spectral bands and index-only inputs) was first conducted using Attention U-Net, followed by a comparative evaluation of representative convolutional and transformer-based architectures, including ResNet-UNet-50, ConvNeXt-UNet, DeepLabV3+ (ResNet-50), and DINOv2-S/14. Model performance is evaluated on an internal Serbian test split (Test SR), an external Serbian dataset (Ext SR), and a cross-regional Balkan dataset (Ext WB). Results: The results demonstrate that compact multispectral configurations (6–9 bands) provide the most stable performance, achieving mIoU values of approximately 0.72–0.74 under in-domain evaluation and remaining robust under external testing. The inclusion of near-infrared and shortwave infrared bands proved critical for effective land cover discrimination, whereas increasing spectral dimensionality beyond this range did not yield systematic improvements in external robustness. Notably, the magnitude of performance degradation under pronounced geographic domain shift exceeds the performance differences observed between architectures under in-domain conditions, indicating that distribution shift exerts a stronger influence on segmentation accuracy than model choice alone. Class-wise analysis revealed agricultural areas as the most domain-sensitive category, while Shapley-based explainability analysis provides additional insight into class-specific spectral dependencies and their role in generalization behavior. Conclusions: Although transformer-based models demonstrated competitive robustness, attention-enhanced convolutional architectures achieved comparable stability across evaluation scenarios. Overall, the findings emphasize the importance of balanced spectral design, class-aware robustness analysis, and explicit out-of-domain evaluation for developing transferable land cover segmentation models in remote sensing applications.
Mitić et al. (Sat,) studied this question.