Random forest classifications for landuse mapping to assess rapid flood damage using Sentinel-1 and Sentinel-2 data

Devastating flood events in northeast Bangladesh are almost a regular phenomenon and last for an abnormally long time for a variety of reasons, including the mountainous topography and the abundance of extreme rainfall events upstream of this region, the complex interaction of the existing river network and haors (wetlands with shallow depressions) in this region. Apart from the monsoon floods, this region frequently experiences flooding during the pre-monsoon (Mar–May) season, known as flash floods, which are considered one of the worst hazards. For conducting the recovery processes after any flood event, accurate damage assessment needs to be done as much as it is possible, which may ensure proper planning and decision-making. Hence, this study proposed a methodology to combine Optical and Radar Sentinel products for flood monitoring and damage assessment. The major advantage of using two different Sentinel products combinedly was while Sentinel-2 could be used for land cover mapping, Sentinel-1 could be used for flood inundation mapping during monsoon periods due to its cloud-penetrating radar imaging property. Gowainghat, an upazila of Sylhet district in Bangladesh, was selected as the case study area, which is prone to monsoons and flash floods almost every year. This area is also one of the most ecological hotspots in Bangladesh, dominated by different haors and beels. Being a dominant waterbody region, it is also necessary to classify the land covers distinctly. A random forest classifier has been introduced in this study to better represent flood extent in this region by ensuring better accuracy of the land cover classification. The results from this study revealed that better results were obtained during land cover mapping when the random forest classifier was applied (90% accuracy) compared to the maximum likelihood classification (74% accuracy). Moreover, the flood damage was evaluated for different land classes providing 89.06% overall accuracies when compared with current land cover. It has been found that about 23.98% of the total agricultural land went underwater during the last monsoon flood, while 72.41% was affected during the last flash flood. The study assisted in determining the losses related to floods and focused on the importance of water management for the target locality.