Implications of Batch Normalization and Dropout on Land Cover Categories in Remote Sensing

Land cover change detection and semantic segmentation are extensively studied research areas in the field of remote sensing. The identification of changes in multi-temporal images and the extraction of semantic segmentation features are crucial. Deep learning techniques are widely used for effective feature extraction due to their automatic feature extraction capability. Convolutional neural networks have been proven to be highly effective for image feature extraction. When extracting features for land cover images, the model may overfit due to class imbalance problems that commonly occur. Some methods utilize batch normalization and dropout to mitigate this issue, but they may not be suitable for all land cover types. Therefore, the proposed work provides a comprehensive analysis of batch normalization and dropout usage based on land cover types. While batch normalization is commonly used in convolutional neural networks, it may not be a good choice for building change detection, as indicated by the proposed analysis. Additionally, the paper explains the usage of dropout not only to reduce the divergence of training and validation data in loss curves but also in building category change detection.