<![CDATA[Precision agriculture increasingly depends on reliable interpretation of high-resolution multispectral satellite imagery to monitor crop variability, detect stress, and support site-specific management. This study aimed to evaluate deep-learning architectures for crop-stress classification and determine whether spectral-spatial integration improves predictive robustness and agronomic interpretability. A quantitative experimental design compared a conventional convolutional neural network, ResNet, Vision Transformer, and hybrid CNN-Transformer using parcel-level training, validation, internal testing, and external spatial-temporal holdout data. Spectral ablation and explainable artificial intelligence analyses were also conducted to assess band importance and model attention. The hybrid CNN-Transformer achieved the highest internal accuracy, macro F1-score, and AUC, while maintaining the strongest performance under external domain shift. Near-infrared and red-edge bands contributed most substantially to classification, particularly for distinguishing intermediate crop-stress conditions. Explainability analysis showed that the best-performing model focused predominantly on canopy regions and vegetation-sensitive spectral information rather than irrelevant background features. The study concludes that reliable precision-agriculture intelligence requires more than high internal accuracy. Integrating local spatial features, broader contextual relationships, multispectral information, external validation, and interpretable model behavior provides a stronger foundation for transferable satellite-based crop monitoring and decision support across heterogeneous agricultural environments. This framework can improve timely intervention, resource allocation, and field-level management decisions.]]>
