Satellite Image Segmentation via Image Quality Enhancement and Modified Unet Architecture

Satellite imaging is essential for urban planning and crisis management as it records detailed spatial information such as buildings, roads, and different types of land cover. Satellite image segmentation is necessary to analyze pictures and extract particular information for applications including urban planning, vegetation monitoring, and time series data analysis due to their complexity. Most segmentation models need input images of size (512,512), which may lead to pixelation or quality reduction due to cropping or zooming. High resolution input data might provide challenges in multiclass segmentation due to its extensive coverage area. This paper present a remodeled Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) to enhance picture quality and maintain crucial data for precise segmentation tasks. The segmentation process involves utilizing a U-Net architecture integrated with a ResNet50 encoder. This decision was taken to utilise ResNet50's advanced feature extraction skills while preserving U-Net's efficacy in obtaining contextual information for accurate segmentation. The network incorporates the Tversky loss function to effectively adjust the influence of various classes in the input data throughout the training process. The model demonstrates an accuracy of 0.94 and a mean Intersection over Union (IOU) of 0.83 in multiclass segmentation tasks, properly outlining different land cover classes such as houses, woodlands, water bodies, and roads.