Detecting sparse building change with ambiguous label using Siamese full-scale connected network and instance augmentation

Building change detection (CD) is an important remote sensing technology with abundant application scenarios. However, the sparsity of building samples in a dataset and ambiguity of labels caused by manual annotations greatly affect the advancement of deep learning (DL) in the field of CD. In this article, we address these problems using two strategies: model optimization and dataset augmentation. On the one hand, we propose a Siamese full-scale connected network based on a feature difference augmentation module (FAM), namely, SFNet. SFNet utilizes full-scale skip connections and deep supervision combined with multiscale feature information to fully capture the features of sparse buildings. Furthermore, we propose a hybrid loss function to alleviate the negative effects of sparse samples and ambiguous labels and use an automatic weight module (AWM) to automatically adjust their weights. On the other hand, we select suitable locations on the original dataset images and blend the virtual buildings harmoniously through the proposed hybrid image blending module (IBM), which further alleviates the sample sparsity problem. Experiments on the LEVIR-CD and WHU-CD datasets show that the proposed method can significantly reduce the impacts of sample sparsity and label inaccuracies and outperform several other state-of-the-art (SOTA) methods.