ABSTRACT
The synthetic aperture radar (SAR) sensors are sensitive to parameters such as scattering mechanism, surface roughness, the dielectric of the target feature, and incident angles. The usability of C-band Sentinel 1 satellite with co-polarized (VV) and Cross-polarized (VH) data proved to be sufficient for accurate flood monitoring. The focus of this paper is to discuss the utility of texture analysis as a powerful input for random forest classifiers to demonstrate the application of SAR data in dynamic flood mapping.
Sentinel 1 interferometric Wide (1W) mode Level-1 ground range detected (GRD) product is pre- processed for this research. A refined lee filter was used for speckle filtering to preserves the texture information and edges. The geocoding was carried out by the Range Doppler orthorectification method using the Shuttle Radar Topographic Mission (SRTM) 3 Sec digital elevation model (DEM) and bilinear interpolation method for resampling. For this study, 11 GRD products consisting of a pre- event, flood, and post-flood scenes for the year 2019 were processed to identify the dynamic flood extent of the Kosi River in Bihar, India. In the present study, Grey Level Co-occurrence Matrix (GLCM) stack proved to be a suitable input dataset for supervised random forest classification.
The proposed approach demonstrates that a random forest classifier is a robust ensemble technique and gives satisfactory results for flood inundation mapping using microwave remote sensing data. The accuracy assessment, performed for all the classified images resulted in an overall accuracy was found to be greater than 88.96 %. The average cohen’s kappa coefficient is 0.82, which indicates substantial agreement. The study also iterates that the adopted methodology can be useful for disaster management applications. However, since the microwave satellites are sensitive to the dielectric of soil, the overestimation of flood extent due to surrounding saturated/waterlogged soil has to be identified rigorously for accurate results. The study concludes that the variation of GLCM descriptors as input feature bands could affect the classification differently, which is primarily associated with the variation of the backscatter throughout the year. In addition, the availability of a high spatial and temporal resolution SAR sensor would better map the dynamics of disaster events like floods, forest fire.
