Advancing Computer Vision: Deep Learning Techniques for Semantic Segmentation in Satellite Imagery

Abstract

A major challenge in computer vision is semantic segmentation, which entails giving labels to specific pixels in an image in order to comprehend fine-grained scene properties. By extracting hierarchical characteristics, deep learning techniques, in particular convolutional neural networks (CNNs), have revolutionized the industry. For many purposes, including urban planning, environmental monitoring, and catastrophe management, satellite imagery interpretation is essential. However, issues like optimization and a lack of labelled training data still exist. In this overview of the literature, we address developments in completely supervised, weakly supervised, and semi-supervised algorithms for semantic segmentation. To increase segmentation accuracy, the suggested methodology fuses the U-Net architecture with GAN-based data augmentation. Comparative findings show that the proposed technique is superior in terms of accuracy, IoU, and DSC. The potential for improving computer vision and the interpretation of satellite images is very high.