In remote sensing, disparity calculation using stereo images is a very necessary task and provides information for estimating the terrain elevation. The fields using disparity of stereo satellite images are used in various fields such as terrain models, autonomous driving using 3D maps, and content development. However, extracting disparity from stereo satellite images is a very difficult task, and inaccurate disparity may be extracted due to complex environments, façade areas of buildings, and texture-less areas. Our proposed method improves feature extraction and 3D aggregation steps based on Gwc-Net using stereo images rectified through RPC (Rational Polynomial Coefficients). To this achieve, we first improve the accuracy of the initial cost volume by extracting important features using the attention module 2D CBAM. In addition, in the aggregation step, we use 3D CBAM to extract important features from the cost volume and use GCE (Correlate-and-Excite) to guide image features to the cost volume to improve disparity. To evaluate the proposed method, the accuracy of disparity is evaluated using RPC-corrected stereo satellite images of DFC2019 data track2 of the US3D dataset. As a result of the experiment, the proposed method exhibited improvement compared to the baseline Gwc-Net.