Abstract
Accurate and transparent classification of agricultural land cover from satellite imagery is essential for operational crop monitoring. This study presents MS-AttNet, a compact convolutional network with squeeze-and-excitation channel attention for Sentinel-2 imagery using all thirteen spectral bands. On the EuroSAT benchmark, evaluated with a stratified 70/15/15 split across five random seeds, the model achieves 98.72 ± 0.21% accuracy, a macro F1-score of 0.9868 ± 0.0021, and a mean one-vs-rest AUC of 0.9999 with 1.32 million parameters. A ResNet-50 baseline re-implemented on the identical split achieves 97.90% accuracy with 23.56 million parameters, making the proposed model approximately eighteen times smaller and nearly three times faster at inference. A controlled ablation using five seeds per configuration shows that squeeze-and-excitation modules, convolutional stage depth, and vegetation indices as additional input channels each change accuracy by less than run-to-run variability. This indicates that the benchmark is saturated with respect to architectural choices in this model family; therefore, performance is not attributed specifically to the attention mechanism. Interpretability is assessed using Grad-CAM and SHAP band attributions with quantitative validation. Against a spatially smooth random control, Grad-CAM reduces deletion AUC from 0.857 to 0.738 and increases insertion AUC from 0.864 to 0.916 across 300 test patches. SHAP identifies blue, near-infrared, green, red, and short-wave-infrared bands as dominant contributors, with stable ordering across independently trained models. NDVI and NDWI were used only for spectral-signature analysis and interpretation, not as classifier inputs. The combination of compact size, competitive accuracy, and quantitatively verified explanations supports large-area operational monitoring applications