Extracting water-related features using reflectance data and principal component analysis of Landsat images

This study aimed to map water features using a Landsat image rather than traditional land cover. We involved the original bands, spectral indices and principal components (PCs) of a principal component analysis (PCA) as input data, and performed random forest (RF) and support vector machine (SVM) classification with water, saturated soil and non-water categories. The aim was to compare the efficiency of the results based on various input data. Original bands provided 93% overall accuracy (OA) and bands 4–5–7 were the most informative in this analysis. Except for MNDWI (modified normalized differenced water index, with 98% OA), the performance of all water indices was between 60 and 70% (OA). The PCA-based approach conducted on the original bands resulted in the most accurate identification of all classes (with only 1% error in the case of water bodies). We therefore show that both water bodies and saturated soils can be identified successfully using this approach.