Open water detection in urban environments using high spatial resolution remote sensing imagery

Abstract

Commonly applied water indices such as the normalized difference water index (NDWI) and the modified normalized difference water index (MNDWI) were originally conceived for medium spatial resolution remote sensing images. In recent decades, high spatial resolution imagery has shown considerable potential for deriving accurate land cover maps of urban environments. Applying traditional water indices directly on this type of data, however, leads to severe misclassifications as there are many materials in urban areas that are confused with water. Furthermore, threshold parameters must generally be fine-tuned to obtain optimal results. In this paper, we propose a new open surface water detection method for urbanized areas. We suggest using inequality constraints as well as physical magnitude constraints to identify water from urban scenes. Our experimental results on spectral libraries and real high spatial resolution remote sensing images demonstrate that by using a set of suggested fixed threshold values, the proposed method outperforms or obtains comparable results with algorithms based on traditional water indices that need to be fine-tuned to obtain optimal results. When applied to the ASTER and ECOSTRESS spectral libraries, our method identified 3677 out of 3695 non-water spectra. By contrast, NDWI and MNDWI only identified 2934 and 2918 spectra. Results on three real hyperspectral images demonstrated that the proposed method successfully identified normal water bodies, meso-eutrophic water bodies, and most of the muddy water bodies in the scenes with F-measure values of 0.91, 0.94 and 0.82 for the three scenes. For surface glint and hyper-eutrophic water, our method was not as effective as could be expected. We observed that the commonly used threshold value of 0 for NDWI and MNDWI results in greater levels of confusion, with F-measures of 0.83, 0.64 and 0.64 (NDWI) and 0.77, 0.63 and 0.59 (MNDWI). The proposed method also achieves higher precision than the untuned NDWI and MNDWI with the same recall values. Next to numerical performance, the proposed method is also physically justified, easy-to implement, and computationally efficient, which suggests that it has potential to be applied in large scale water detection problem.