Abstract
Abstract. Efforts on flood mapping from active and passive satellite Earth Observation sensors increased in the last decade especially due to the availability of free datasets from European Space Agency’s Sentinel-1 and Sentinel-2 platforms. Regular data acquisition scheme also allows observing areas prone to natural hazards with a small temporal interval (within a week). Thus, before and after datasets are often available for detecting surface changes caused by flooding. This study investigates the contribution of textural variables to the predictive performance of a data-driven machine learning algorithm for detecting the effects of a flooding caused by Sardoba Dam break in Uzbekistan. In addition to the spectral channels of Sentinel-2 and polarization bands of Sentinel-1, two spectral indices (normalized difference vegetation index and modified normalized difference water index), and textural features of gray-level co-occurrence matrix (GLCM) were used with the Random Forest. Due to high dimensionality of input variables, principal component (PC) analysis was applied to the GLCM features and only the most significant PCs were used for modeling. The feature stacks used for learning were derived from both pre- and post-event Sentinel-1 and Sentinel-2 images. The models were validated through model test measures and external reference data obtained from PlanetScope imagery. The results show that the GLCM features improve the classification of flooded areas (from 82% to 93%) and flooded vegetation (from 17% to 78%) expressed in user’s accuracy. As an outcome of the study, the use of textural features is recommended for accurate mapping of flooded areas and flooded vegetation.