Enhancing DeepLabV3+ for aerial image semantic segmentation using weighted upsampling

Segmentation of land cover in aerial images is a major challenge that has led to the proposal of various solutions. Among these, the DeepLabV3+ architecture appears to be one of the most promising approaches. However, despite its good results for segmentation, there is still a need to increase its robustness and improve its performance. One interesting approach is to work on the components of this architecture by integrating new operational blocks or modifying certain internal processes. In this paper, we introduce a refinement of the DeepLabV3+ architecture in which we enhance the upsampling method in the decoder and added some extra layers of convolutions to improve aerial image semantic segmentation. As a case study, we employed the LandCover.ai (Land Cover from Aerial Imagery) dataset and achieved a mean Intersection over Union (mIoU) of 83.49% without data augmentation and an improved mIoU of 85.68% with data augmentation.