Given the high cost and difficulty of shadow annotation, the shadow detection performance of supervised learning models in high-resolution remote sensing images is severely limited mainly because of the lack of training samples. To solve this problem, this study proposes a shadow detection method that is based on shadow direction prior, which has unique advantages in learning the inherent shadow information contained in unlabeled high-resolution remote sensing images. Through self-supervised learning methods, the advanced semantic information of ground object shadows contained in remote sensing images can be deeply explored, thereby improving the applicability and accuracy of the model. This innovative approach provides a new way to overcome the data limitation problem in shadow detection and offers new possibilities for improving the model’s ability to understand complex high-resolution images.This study proposes an innovative shadow detection method for high-resolution remote sensing images that is based on prior knowledge of the shadow direction and explores the effectiveness of shadow direction prior in representing high-level semantic features of object shadows in remote sensing images. The key task in this method is to construct an auxiliary task through shadow direction prior so that the deep neural network can effectively understand the shadow of objects in remote sensing images. Moreover, to further enhance the deep neural network’s ability to learn the key features of object shadows, this study designs a direction transformation-independent noise processing mechanism and data enhancement strategy for self-supervised shadow detection, both of which improve the generalization ability of the model.The effectiveness of the proposed method is proven through verification on the Aerial Imagery Dataset for Shadow Detection (AISD). Experimental results show that compared with the traditional Unet benchmark model, the proposed method can successfully learn prior knowledge in unlabeled high-resolution remote sensing images. By leveraging this prior knowledge, the developed method improves the accuracy of the Unet model and demonstrates competitive efficiency and performance in small sample cases. The results of this study highlight the superiority of the proposed method in addressing the challenge of shadow identification in high-resolution remote sensing images, especially in the absence of labeled samples. By fully utilizing prior knowledge, the method successfully extends the performance of the Unet model and provides a promising solution for processing unlabeled high-resolution remote sensing images in practical applications.The qualitative and quantitative results consistently show that using a shadow direction prior method can considerably improve the accuracy of shadow detection. The use of the DTICT strategy helps alleviate the interference caused by shadow-free samples and samples containing shadows in special directions during network training. The data enhancement strategy combined with contrast cropping and color transformation provides the network with diverse training samples, thereby enhancing the pre-trained model’s ability to extract object shadow features. The proposed method achieves competitive object shadow detection accuracy by using a small number of labeled samples.

shadow detection;self-supervised learning;data augmentation;shadow direction prior;high-resolution remote sensing imagery