Multi-Feature Dynamic Fusion Cross-Domain Scene Classification Model Based on Lie Group Space

Abstract

To address the problem of the expensive and time-consuming annotation of high-resolution remote sensing images (HRRSIs), scholars have proposed cross-domain scene classification models, which can utilize learned knowledge to classify unlabeled data samples. Due to the significant distribution difference between a source domain (training sample set) and a target domain (test sample set), scholars have proposed domain adaptation models based on deep learning to reduce the above differences. However, the existing models have the following shortcomings: (1) insufficient learning of feature information, resulting in feature loss and restricting the spatial extent of domain-invariant features; (2) models easily focus on background feature information, resulting in negative transfer; (3) the relationship between the marginal distribution and the conditional distribution is not fully considered, and the weight parameters between them are manually set, which is time-consuming and may fall into local optimum. To address the above problems, this study proposes a novel remote sensing cross-domain scene classification model based on Lie group spatial attention and adaptive multi-feature distribution. Concretely, the model first introduces Lie group feature learning and maps the samples to the Lie group manifold space. By learning features of different levels and different scales and feature fusion, richer features are obtained, and the spatial scope of domain-invariant features is expanded. In addition, we also design an attention mechanism based on dynamic feature fusion alignment, which effectively enhances the weight of key regions and dynamically balances the importance between marginal and conditional distributions. Extensive experiments are conducted on three publicly available and challenging datasets, and the experimental results show the advantages of our proposed method over other state-of-the-art deep domain adaptation methods.