Satellite Image Segmentation Based on Improved Lightweight U-Net

Accurate segmentation of satellite images is crucial for precise environmental monitoring, resource management, and disaster response. The data foundation primarily relies on images from satellites, analyzed through deep learning techniques. However, there is still a need to enhance recognition accuracy and reduce the model size. To address this issue, the paper proposes an Improved Lightweight U-Net. The objective is to segment satellite images using fewer layers and parameters, which will help reduce storage requirements, computational complexity, and training time. The input image is resized to 64x64 pixels. The encoder consists of two downsampling layers with convolutional and max-pooling operations to extract features, followed by the bottleneck layer that processes the extracted features. In the decoder, two upsampling layers restore spatial dimensions, with skip connections from the encoder to refine the output. The final output layer applies a 1x1 convolution with a sigmoid activation function to generate the segmentation map. The proposed Improved Lightweight U-Net achieves an accuracy of 99.23%, an IoU of 97.33%, and an F1 score of 98.58%.