Publication

Abstract : Experts in remote sensing have focused a lot of effort on the extraction of geographical objects from high-resolution aerial data using supervised semantic segmentation and deep neural networks. Unfortunately, the extraction procedure frequently produces imperfect results because of the complex structure of geographical information, sensing resolution constraints, and scene occlusions. In this study, we propose the integration of a generative model to post-process deep neural network segmentation outputs. To achieve higher computational efficiency, we specifically use a conditional Generative Adversarial Network that is based on Pix2pix. When paired with basic pre-processing methods like edge-preserving filtering, our suggested strategy shows significant improvements in road network segmentation over current approaches. Trial assessments carried out on the Massachusetts road image dataset provide encouraging outcomes, as the suggested method attains an precission of 81.9% and recall of 69.3%. These findings translate into an F1-score of 91.2%, and a Intersection over Union (IOU) of 57.4%. Comparative studies show that our GAN-based framework outperforms conventional CNN-based methods, especially when it comes to edge information preservation.