Comparative evaluation of spectral water indices for monitoring temporary ponds in the Ferlo region (Senegal)

In the context of increasing climate variability, accurately monitoring the availability of water in temporary ponds in the sylvopastoral zone of Ferlo (Senegal) is crucial for pastoral resilience and food security. This study evaluates the performance of six spectral indices derived from Sentinel-2 imagery, Modified Normalized Difference Water Index (MNDWI), Normalized Difference Water Index (NDWI), Normalized Difference Pond Index, Tasseled Cap Wetness (TCW), Normalized Difference Moisture Index, and Automated Water Extraction Index shadow version (AWEI_sh), for surface water detection, using ultra-high-resolution drone-based orthophotos (3 cm) as reference data. Ten ponds (eight natural and two artificial) were selected. The orthophotos enabled a detailed classification of Sentinel-2 pixels into three land cover classes (water, bare soil, and vegetation) and allowed the calculation of the proportion of each class at the pixel level. The indices were first calibrated using receiver operating characteristic curves, the minimum distance ( D min ) to the perfect classification point across various water coverage thresholds (2% to 100%). Their performances were then evaluated and compared using confusion matrix metrics and McNemar’s test with the validation dataset. The results highlight three particularly effective indices: (i) AWEI_sh provided the best overall trade-off between completeness and precision; (ii) TCW showed balanced performance across all cover ranges and was preferable when minimizing false positives is critical; (iii) MNDWI performed best under high water coverage conditions. A logistic regression applied to misclassified pixels revealed that vegetation significantly increases the likelihood of false positives (odds ratio > 1; p < 0.005), whereas bare soil had a more moderate effect. These findings underscore the influence of non- water classes on classification errors and highlight the importance of integrating vegetation indices into future water detection models. Study limitations include potential limited reliability of index thresholds and generalizability due to differing sensor platforms used temporally limited baseline constrained by field observations. By combining ultra-high-resolution drone data, rigorous spectral index validation, and statistical modeling of classification errors, this study offers a robust approach for fine-scale mapping of temporary water bodies in the Sahel. It provides a valuable tool for the adaptive and sustainable management of water resources in pastoral systems.