Semantic Segmentation of Satellite Images using Modified U-Net

Detecting roads, regions, vegetation-flora, and evidence of water resources in regions is essential for the long-term development and enhancement of remote areas around the world. Despite the fact that deep neural networks have made tremendous progress in the semantic segmentation of satellite pictures, the majority of present techniques yield unsatisfactory results. The challenge of retaining the quality of semantic segmented pictures is addressed in this study by presenting a unique combination of architecture. The segmentation model offers a solution for generating automatic area segmentation and shows high accuracy for six classes: Building, land, road, vegetation, water, and miscellaneous. The model is trained on Dubai Satellite imagery dataset of MBRSC. For achieving the best performance we augmented the dataset and trained the augmented data where Shift, scale, and rotate transformations were applied to the satellite images and segmentation masks. This baseline U net model modifies U-Net's encoder by employing the Inception ResNet V2 model to have enhanced mathematical and structural complexity. The model is further evaluated using the Dice coefficient and pixel accuracy which came out to be 82 percent and 87 percent respectively in validation samples.