A comparative study of different deep learning models for land use and land cover mapping of flood detention basin

Abstract

Abstract Flood detention basin (FDB) is an important part of the flood control system in the basin. It is of great significance for scientific flood control to obtain the land use and land cover (LULC) classification map of FDB with higher accuracy quickly and accurately. In recent years, deep learning has shown great potential in LULC classification. In this study, we make LULC training dataset and explore three state-of-the-art (SOTA) DL architectures: Unet++, ResUnet++, DeepLab v3+ across Mengwa FDB. The experiments show all methods used in this study are available for LULC classification, which overall accuracy is ResUnet++(95.11%), Unet++(91.92%), DeepLab v3+ (91.04%) and kappa coefficient is ResUnet++(0.91), Unet++(0.85), DeepLab v3+ (0.83). Deep learning methods can be used for automatic extraction of LULC from high-resolution satellite images, providing data support for economic loss assessment and post disaster construction in flood storage areas.