Performance Comparison of Convolutional Neural Network Deep Learning Architectures for Remote Sensing Image Segmentation

In this paper, the performance of five convolutional neural network (CNN) deep learning architectures were evaluated for delineating high complexity building regions in remote sensing (RS) images. They are the U-Net, SegNet, ResU-Net, AttentionU-Net, and V-Net architectures that are commonly used for image segmentation. In the existing works, the robustness of these deep learning architectures was evaluated in delineating the regions of interest (ROIs) in various RS image as well as segmenting anatomical structures as ROIs in the medical images, individually or in pairs. It is crucial to evaluate the performance efficiency of these architectures on the same image dataset that has high complexity of ROIs for delineation. Therefore, the performances of the five mentioned deep learning architectures were evaluated for comparison in segmenting buildings regions as ROIs in the WHU RS images dataset, which consist of various colors, shapes, sizes, and textures features. The experimental results showed that the AttentionU-Net deep learning architecture accurately segmented the buildings regions achieving an average Intersection over Union (IoU) of 0.9780, 0.9436, 0.9469 and average F-measure of 0.9690, 0.9230, 0.9301 for training, validation, and test images, respectively with 62 epochs in comparison to the other four architectures. This is mainly because of the incorporation of the attention gates (AGs) in the AttentionU-Net architecture that are able to extract the fine feature details of the complex building regions in the WHU RS images dataset.