A hierarchical self-attention augmented Laplacian pyramid expanding network for change detection in high-resolution remote sensing images

Change detection methods can achieve high learning ability and recognition accuracy with the introduction of deep convolutional neural networks, but due to the influence of the convolution kernel and deep feature sampling, problems such as the limited feature extraction range and loss of information are inevitable. In addition, there can be severe imbalance in the proportions of changed and unchanged pixels compromising the stability of model training. To address these interrelated problems, we propose a hierarchical self-attention augmented Laplacian pyramid expanding network (H-SALENet) for supervised change detection in high-resolution remote sensing images. H-SALENet is composed of an encoder and a decoder. In the encoder, H-SALENet combines a deep convolutional module with a hierarchical and long-range context augmentation module (HLAM) to extract the deep features of bi-temporal images; the representation capability of multi-level and long-range dependent change features is enhanced through deep convolution and 2D transformer-structured multihead self-attention learning. In the decoder, a Laplacian pyramid expansion module (LPEM) is proposed to catch change information at different scales and reduce high-frequency information loss simultaneously, thus weakening the influence of deep feature resampling on the change map. In addition, a data-balanced loss function concentrating on both the quantity and the complexity of the pixels was designed to reduce the influence of the imbalance between changed and unchanged pixels. H-SALENet was tested on two kinds of public datasets; the qualitative and quantitative experimental results show that the proposed network outperformed three benchmark change detection networks in terms of the integrity of change objects and the capability to obtain change contours.