A Global-local Feature Interaction Network for Anti-cloud Interference Change Detection based on Contrastive Learning

Bi-temporal change detection (CD) constitutes a vital task within the domain of remote sensing (RS). due to pseudochange caused by factors such as seasonal climate variations, temporary objects, and cloud interference. To address the pseudochange issue stemming from cloud interference, we propose a global-local feature interaction anti-cloud interference change detection network (ACCDNet) based on contrastive learning. To mitigate the scarcity of remote sensing datasets with cloud interference, we leverage contrastive learning principles, using remote sensing images of the same scene at the same time with and without cloud interference as positive sample pairs to pre-train anti-cloud interference capability of the feature extractor. The proposed CD network which bases on a CNN-transformer network, utilizes Resnet to extract multi-scale features from the original input images. We introduce transformer modules and attention mechanisms to effectively extract the contextual information within the input images. To address the pseudochange issue caused by cloud interference, we simulate the effect of thin cloud interference using Perlin noise and add it to classical datasets. Experimental results on LEVIR-CD and LEVIR-CD-cloud datasets augmented with cloud interference demonstrate the priority and efficiency of the proposed method.