Enhancing the classification of spectrally similar land use/land cover classes using transfer learning in arid regions

Accurate mapping of Land Use/ Land Cover (LULC) in arid and semi-arid areas is essential for managing urban planning, monitoring environmental risks, and preserving the ecosystem balance. Conventional classification techniques provide inaccurate mapping of LULC in such areas due to the spectral similarity behavior of built-up and bareland classes. To address this challenge, semantic segmentation techniques have been recently employed to enhance classification accuracy. Meanwhile, in the case of imbalanced datasets, these models still exhibit limited generalization ability. To overcome this limitation, this study proposes a transfer learning approach to enhance the classification accuracy between built-up and bareland classes in the Nile Delta in Egypt, utilizing an imbalanced dataset. Four transfer learning models were employed, including Resnet50-Unet, Resnet50-FPN, Resnet50-PSPNet and Unet and their results were compared with the Maximum Likelihood classifier. The results showed that Resnet50-FPN achieved the highest F1-score of (0.877), followed by Resnet50-Unet (0.8705), Resnet50-PSPNet (0.852), Unet++ (0.792) and Maximum Likelihood (0.725). These findings highlight the effectiveness of transfer learning in enhancing LULC classification performance compared to the conventional Maximum Likelihood classifier. The superior performance of Resnet50-FPN is due to its multi-scale feature extraction capability, which enables more effective representation of LULC classes while preserving the spatial information of remotely sensed images. This indicates that the ResNet50-FPN classifier can be employed in different contexts to differentiate between built-up and bareland areas.