Evaluation of typhoon-induced inundation losses associated with LULC using multi-temporal SAR and optical images

Abstract

We utilized a two-branch end-to-end network (MultiSenCNN) for land use and land cover (LULC) classification and flood event mapping using multispectral (MS), panchromatic (Pan) and synthetic aperture radar (SAR) images, where flooding was induced by typhoon Lekima in August 2019. Flood damages were assessed by considering both the LULC and flood maps. We defined three strategies to compare the MS + SAR and MS + Pan images to demonstrate the ability of the MultiSenCNN algorithm for LULC classification. The three strategies yielded an average overall accuracy of ∼98% and an average Kappa of ∼0.98 for LULC classification. The overall accuracy of the fused MS + SAR images is slightly higher than the MS + Pan images when using the same model training samples. The flood mapping shows an overall accuracy of 97.22% and a Kappa of 0.94, with a flood inundation area of 101 km2 that mainly inundated cropland and urban areas. Compared to other LULC types, the flooded cropland has caused more loss of ecosystem service values during typhoon Lekima, accounting for 81.19% of the total. Using SAR mages can well monitor the start/end states of flood events and the inundated areas, providing the flood status information to rescuers and governments for making timely decisions.

Keywords

• LULC classification
• flood mapping
• synthetic aperture radar
• deep learning
• inundation assessment