Current remote sensing target detection methods based on deep learning can only identify the type and location of remote sensing targets, but cannot detect the spatial relationship between remote sensing targets. Aiming at this problem, a method for detecting the spatial relationship of remote sensing targets is proposed. First, a convolutional neural network is used to construct a vision module to extract the visual features in the remote sensing image. Second, a semantic module is constructed to map the extracted visual features to the semantic embedding space to achieve the deep fusion of the visual features and semantic features of the remote sensing target. Finally, the Softmax function and the visual consistency loss function are introduced into the traditional triplet loss function, and an improved triplet loss function is designed. The proposed method is used to conduct experiments on the self-made remote sensing target spatial relationship detection dataset. The experimental results show that among the top 20, 50 and 100 prediction results, the recall rates of the proposed method are 76.32%, 78.54% and 81.47%, respectively, indicating that the proposed method has good spatial relationship detection performance and can accurately detect remote sensing objects and their spatial relationships in remote sensing images.