Abstract
The automatic segmentation of water bodies from remote-sensing satellite images offers valuable insights into water resource management, flood monitoring, environmental changes, and urban development. However, extracting water bodies from satellite imagery can be challenging due to factors such as varying water body shapes, diverse environmental conditions, cloud cover, and shadows. These difficulties have a significant impact on waterbody segmentation, particularly in precisely maintaining high-quality segmented images and determining the boundaries of waterbodies. To overcome these issues, researchers have introduced several approaches; however, they suffer from precisely identifying the boundaries of waterbodies due to their irregular shapes. This difficulty is particularly pronounced in traditional threshold-based and machine-learning techniques, which often struggle to achieve accurate segmentation when confronted with complex structures, cluttered backgrounds, or objects of varying sizes and shapes. The objective of this research is to develop innovative deep-learning (DL) approaches to address these challenges and enhance the accuracy of waterbody segmentation in Remote sensing applications. This research introduces a deep learning model, namely AER U-Net architecture, which integrates advanced architectural elements into U-Net, such as residual blocks, self-attention mechanisms, and dropout layer, due to which the model significantly enhances segmentation accuracy and generalization capability. The architecture employs a contracting path consisting of convolutional layers, batch normalization, and activation layers to extract multi-scale features. Residual blocks improve feature learning efficiency while addressing the vanishing gradient issue through the inclusion of skip connections. Dropout layers in the encoder and bottleneck paths are incorporated for regularization, reducing the risk of overfitting. Additionally, the attention mechanism ensures precise refinement of skip connections, further improving segmentation performance. The model is trained using the Adam optimizer combined with a binary cross-entropy loss function, making it highly effective for binary segmentation tasks with an IoU score of 0.94, highlighting its effectiveness for practical environmental applications.