Multiscale Change Detection Network Based on Channel Attention and Fully Convolutional BiLSTM for Medium-Resolution Remote Sensing Imagery

Remote sensing change detection (CD) is used to detect the difference in the state of objects or phenomena by observing it at different times. CD is widely used in disaster monitoring, land-use and land-cover change analysis, urban expansion detection, and other fields. Medium-resolution (MR) remote sensing imagery can be used for global and regional CD due to the real-time acquisition, extensive coverage, and historical data advantages. Therefore, medium-resolution remote sensing imagery change detection (MRCD) is a very important topic. Compared with very high resolution (VHR) imagery, MR imagery has less texture and edge information. Besides, the object has a large-scale size in VHR imagery scene while the same object will only have a small-scale size in MR imagery scene. To solve the challenge of MRCD, we propose a joint spatial–spectral–temporal network for MRCD, named Multiscale Convolution Channel Attention coupling full convolutional BiLSTM Network (MC
²
ABNet). The MC
²
ABNet consists of multiscale convolutional channel attention (MC
²
A) module and fully Convolutional Bidirectional Long Short-Term Memory (ConvBiLSTM) network. In the encoder, MC
²
A module is used to extract multiscale spatial features from multitemporal imagery at each encoding level by sharing structure, parameters, and weights. In each MC
²
A module, the multiscale convolution extracts multiscale spatial features with different receptive fields, and then the channel attention is used to ease the information redundancy during downsampling. The ConvBiLSTM is applied to calculate the time difference features in both forward and backward directions and utilizes spatial information synergistically to smooth change noise for obtaining complete time difference features. The extensive experiments have been conducted on ONERA satellite change detection and SpaceNet7 datasets. Compared with other state-of-the-art methods, our network achieves the highest accuracy on both datasets.