Progressive Self-Optimization Network: An unsupervised change detection method for VHR optical remote sensing imagery

• An unsupervised method (PSONet) is proposed for VHR imagery change detection. • A sampling strategy eases the trade-off between sample accuracy and quantity. • A lightweight CNN is designed to detect changes using fewer parameters and FLOPs. • A progressive inference pattern optimizes the network stepwise for stable results. Change detection in very-high-resolution (VHR) remote sensing imagery has consistently been a focus and challenge within the remote sensing community. We present a novel unsupervised method named Progressive Self-Optimization Network. In this method, a new sample strategy is developed based on the initial change detection across three feature domains: spectral, deep, and class signal, to capture the “weak-to-strong” change signals and collect training samples with high accuracy. A new lightweight convolutional neural network is designed by partially replacing traditional convolutional layers with weight-shared partial convolution. To detect changes, a progressive self-optimization pattern is proposed based on the “weak-to-strong” change signals. This pattern detects changes gradually from regions with weak and strong change signals to regions with moderate change signals in a progressive manner. During this progressive process, detection results from each progression are integrated with “weak-to-strong” change signals to reselect training samples for the transfer training in the following progression, thus attempting to optimize the lightweight network. The final change map is generated by fusing all progression results. Five open VHR datasets and fourteen state-of-the-art unsupervised methods validate the proposed method.