Urban surface water body detection with suppressed built-up noise based on water indices from Sentinel-2 MSI imagery

Water spectral indices can enhance the difference between water bodies and background features. Thus, they have been widely used to extract and map surface water bodies based on multispectral satellite imagery. The urban scene is very heterogeneous since the surface is composed of a vast diversity of man-made objects, often of mixed distribution. Urban surface water mapping faces an extreme overestimation phenomenon because certain types of objects such as shadow, dark roads and some artificial features may return similar values to water bodies after an index computation. This study proposes a noise-prediction strategy to eliminate such misclassified nonwater areas in an automated way. Constrained energy minimization (CEM), a typical sparse target detection algorithm that does not need any background information, is utilized to draw the possible distribution of noise based on prior noise samples. The initial noise samples are automatically extracted by calculating the difference between two water indices widely accepted in urban scenes, namely, the modified normalized difference water index (MNDWI) and the automated water extraction index (AWEI). Recently freely available Sentinel-2 multispectral satellite imagery, with high spatial resolution (up to 10 m) and high repeated global coverage (every 5 days), was adopted, considering its potential on urban land cover mapping. Compared with the AWEI based approach, the results show that the proposed noise-prediction approach obtained an improved overall accuracy (increased Kappa coefficient by 0.07 on average), dramatically enhanced user accuracy (by 12.47% on average) with reduced noise, and simultaneously slightly decreased producer accuracy (by −1.19% on average). That is, the proposed method possesses an improvement of the misclassification of nonwater bodies to water bodies and a suppression of the missing of water body extraction at the same time. Finally, the comparative results, with the varying water index segmentation thresholds (−0.2 to 0.3) and an automatic Otsu threshold, indicate the robustness to the threshold of the proposed approach. •Map urban surface water bodies and suppress the background in an automated way•Noise prediction via the different water indices and CEM target detection•Balance the misclassification of nonwater bodies and missing of the water bodies•The default threshold of zero achieves near optimal mapping performance