Geospatial Prior Guided Confused-Class Data Augmentation for Semantic Segmentation of Remote Sensing Images

Semantic segmentation of high spatial resolution remote sensing images plays a crucial role in diverse applications. Deep learning-based methods have emerged as the dominant approach in this field. However, existing methods face the challenge of effectively using the information in annotated training data, resulting in limited model performance. Data augmentation is a commonly used technique to better reveal the information embedded in the data through appropriate transformations. However, these methods usually fail to account for the unique characteristics of remote sensing images, such as the high spectral similarity among different classes and intra-class variation among distinct objects. This oversight hinders the model’s ability to distinguish between spectrally similar classes, resulting in persistent misclassifications. To address these challenges, we propose a novel geospatial prior guided confused-class data augmentation (GCA) method for semantic segmentation of remote sensing images. By emphasizing the occurrence of confused classes within augmented training data, GCA compels the model to focus on learning discriminative features to improve class separability. GCA consists of three key modules: geospatial prior information extraction module (moduleG), confused-class similar units detection module (C), and confused-class data augmentation module (moduleA). Specifically, moduleGextracts geographic units and their spatial relationships. ModuleCidentifies highly similar units from confused classes that are prone to misclassification. ModuleAgenerates challenging training data by emphasizing confused-class similar units under the guidance of geospatial priors. Experiments on the GID and LoveDA datasets demonstrate that the proposed GCA method shows a better performance than state-of-the-art data augmentation methods. The GCA code is publicly available at https://github.com/Jingying-Li/GCA.