IA-CDNet: Change Detection in Adverse Remote Sensing Image Conditions With an Advanced Image-Adaptive Method

Change detection using deep learning methods in optical remote sensing imageries has become the mainstream method, primarily due to their strong spatial-temporal transferability. However, current change detection models are designed for high-quality optical remote sensing imageries and seldom account for image conditions such as haze cover or low-light scenarios. As we know, the quality of optical remote sensing imageries can be severely degraded by haze or low-light conditions, which in turn significantly impairs the performance of change detection models. Yet there is still a lack of advanced change detection models designed for the aforementioned adverse scenarios at present. To address this gap, we proposed a simple image-adaptive change detection model named IA-CDNet, aimed at improving model performance in adverse image conditions. Initially, we developed an image-adaptive filter module, which employed several differentiable image filter operators to enhance the quality of the inputted image. Notably, the parameters of these filters are automatically learned using a lightweight convolution neural network (CNN) module. Then, a generic change detection model is integrated to extract low-level and high-level semantic information, facilitating the detection of changes between bi-temporal remote sensing imagery. Ultimately, experiments conducted on three open-sourced change detection datasets and two large-scale real-world scenarios have demonstrated the effectiveness of IA-CDNet. We anticipate this model can make an important contribution to change detection research fields and improve model performance further. The Open sourced code can be accessed at: https://github.com/wzp8023391/Change-detection-in-adverse-conditions.