Beyond accuracy: environment-dependent performance and threshold robustness of spectral water indices across regions and seasons

Spectral water indices are widely used for operational surface-water mapping, yet large-scale automation is often hindered by threshold uncertainty and limited cross-scene transferability. Existing studies typically rank indices by peak accuracy while overlooking the critical trade-off between maximal performance and sensitivity to threshold perturbations. Here we establish a broad multi-region, multi-season benchmark using Sentinel-2 imagery, covering 12 environmentally heterogeneous regions and four seasons (48 scenes). Twelve representative indices are assessed in terms of scene-adaptive peak accuracy (F1max) and threshold robustness, quantified by a newly proposed Robustness Smoothness Index (RSI). We also introduce PCA-Bridge, which embeds scenes using environmental variables only and maps performance and robustness landscapes in the resulting environmental space. Rather than identifying a single universal index, our analysis reveals a clear trade-off between peak accuracy and threshold robustness. While a median F1max>0.92 across 48 scenes suggests generally high performance, performance declines markedly in complex environments, particularly those dominated by snow and ice. NDWI attains high peak accuracy (F1max≥0.95) in 50.0% of scenes but shows low robustness (RSI≥0.75 RSI≥0.75 RSI≥0.75 in only 8.3%), indicating strong threshold sensitivity. By comparison, MBWI and AWEInsh achieve the highest median F1max (0.968 and 0.960), whereas WI2015 and NWI exhibit the strongest robustness overall, with the nhighest mean RSI (0.840 and 0.837) and broad robustness coverage (RSI≥0.75 RSI≥0.75 RSI≥0.75 in 87.5% of scenes). These findings support task-oriented index selection rather than one-size-fits-all ranking. Out-of-scene transfer experiments showed that scene-wise optimal performance does not guarantee threshold generalisation. Overall, the proposed framework helps identify when peak accuracy, robustness, and transfer performance align or diverge, supporting more reliable large-scale surface-water monitoring.