Deep Adaptive Pix-2-Pix Conditional Generative Adversarial Networks for Semantic Segmentation of Medium Resolution Google Earth Imagery

Abstract

Land cover analysis using aerial images is vital for understanding how abiotic and biotic components are clustered on the earth’s surface. This research paper primarily focuses on the semantic segmentation of aerial remote sensing images using proposed adaptive Pix2Pix conditional generative adversarial networks. The network model architecture contains a generator phase with a modified U-Net architecture. A discriminator characterized by a convolutional PatchGAN as a classifier. On the generator side, we replaced the backbone of the standard U-Net feature extractor with two classical deep feature extractors, ResNet34 and InceptionNet-V3. The paper experimentally compares the effect of modified Unets in Pix2Pix Conditional Generative Adversarial Nets for semantic segmentation of aerial imagery for land cover application. A comparison is made between the standard Pix2Pix GAN and the proposed adaptive Pix2Pix cGAN. Google Earth images are used to create the dataset. The experimental results are evaluated and compared using metrics like Root Mean Square Error (RMSE) and Structural Similarity Index (SSIM). The results signify that the proposed adaptive model built using U-Net with Inception-Net: V3 backbone as Generator Pix2Pix cGAN as Discriminator outperforms the standard approach, giving an RMSE of 0.062 and an SSIM of 0.513. The subjective evaluation and objective (visual) analysis both conclude that the proposed adaptive pix2pix cGAN approach is accurate for the semantic segmentation of aerial images.