The problem of unbalanced category distribution of foreground and background in remote sensing change detection (CD) tasks has become a key problem in improving the performance of CD models when processing complex remote sensing image data. To address this issue, we develop a novel adaptive cross-spatial sensing network (ACS-Net), aiming to dynamically adjust the degree of attention of the foreground and background to achieve accurate and efficient CD. The proposed ACS-Net combines the global context module (GCM) and the cross-spatial learning module (CSLM) by introducing the adaptive weight adjustment mechanism (AWAM) to adaptively adjust the weight coefficients of the GCM and CSLM, thereby enhancing the ability to extract foreground features. GCM captures the global context information of the image to generate global perceptual features, thereby helping to distinguish the foreground and background. While CSLM focuses on learning the information association between different spatial locations, which can enhance the model’s sensitivity to spatially changing details, and then focus on local feature extraction of foreground. The experimental results show that the F1-Score index of ACS-Net on two public datasets (Levir-CD and Lebedev-CD) reaches 91.53% and 96.25% respectively, validating its effectiveness and superiority.