Semi-Supervised Semantic Segmentation in Remote Sensing Images Using Active Learning and Deep Global Second Order Pooling Network

Training convolutional neural networks (CNNs) for semi-semantic segmentation in remote sensing images is difficult due to the substantial amount of labelled data required. The scarcity of dense pixel-level annotations hampers the model's ability to extract meaningful information from the images. We propose a novel semi-supervised semantic segmentation network called 3S-Net, designed explicitly for high-resolution remote sensing imagery to overcome this challenge. This network leverages a limited number of labelled examples and many unlabeled examples to generate pseudo-labels, expanding the labelled training set. The Deep Global Second-Order Pooling Network (DG2oPN) is employed to learn discriminative and representative features from the entire image information throughout the network. Our approach involves two main steps: First, an active learning-based sampling strategy is utilized to select a representative set of labelled data. Secondly, we introduce the 3S-Net, which learns from limited annotated data and effectively leverages unlabeled images by utilizing the second-order statistics of the entire image. Experimental results on two public remote sensing datasets demonstrate that our semi-supervised semantic model efficiently and effectively addresses the issue of limited labelled samples, outperforming other state-of-the-art methods.