Spectral indices such as the Normalized Difference Water Index (NDWI), Modified NDWI (MNDWI), and Automated Water Extraction Index (AWEI) encode decades of domain knowledge for water body detection, yet deep learning models rarely exploit these indices beyond post-processing comparisons. We demonstrate that incorporating NDWI, MNDWI, and AWEI as explicit input channels to a DeepLabV3+ architecture achieves 99.17% tile-aggregate F1-score on the S1S2-Water benchmark (per-scene mean: 84.61%), exceeding classical NDWI (97.20%) by 1.97 percentage points on the same tile-aggregate metric. Full-scale ablation studies on all 45 training samples reveal that spectral indices contribute +15.26 percentage points over the 5-channel baseline (83.91% \(\rightarrow \) 99.17%), while spectral consistency loss was counterproductive (−12.37pp), suggesting that input channel design is a more effective integration strategy than loss function constraints for domain knowledge. Individual index ablation confirms that AWEI contributes an additional +0.81pp beyond NDWI and MNDWI alone (98.36% \(\rightarrow \) 99.17%). To validate the architecture’s extensibility, we conducted SAR-optical fusion experiments using Sentinel-1 VV and VH channels alongside the 8-channel optical model. The 10-channel fusion model achieved 93.59% F1; however, 99% of SAR pixels in this dataset are invalid fill values, so this result documents a data quality constraint rather than a validated multi-modal fusion demonstration. Statistical analysis confirms that the training subset is representative of the full distribution (Kolmogorov-Smirnov test, \(\varvec{p > 0.05}\) ). Qualitative comparison across 32 diverse test scenes shows complementary error profiles between the deep learning model and classical NDWI, suggesting potential for ensemble approaches. These results establish spectral index channel engineering as a simple yet highly effective strategy for integrating domain knowledge into deep learning for environmental monitoring.
