PC-PPSF: A Principal Component-Driven Pyramid Perception Segmentation Framework for High-dimensional Remote Sensing Image Semantic Interpretation

The semantic parsing task of high-dimensional remote sensing images faces multiple challenges, including redundant input feature dimensions, severe inter-class spectral structure coupling, and modeling degradation issues caused by spatial heterogeneity. This paper proposes a segmentation framework named principal component feature reconstruction with pyramid context aggregation (PC-PPSF), which integrates principal component-driven feature reconstruction with a pyramid context aggregation mechanism, establishing a unified modeling pathway from input space regularization to enhanced semantic representation. Specifically, linear mapping of the original multispectral input is performed through principal component analysis (PCA) to effectively compress the high-dimensional feature space and enhance inter-class separability and low-dimensional embedding capabilities. At the structural level of the model, the pyramid scene parsing network (PSPNet) is introduced, leveraging its multi-scale pooling structure to achieve deep integration of global and local context information, improving semantic consistency and boundary structure sensitivity. Experimental results show that the proposed PC-PPSF significantly improves segmentation accuracy while maintaining a lightweight network, achieving an average mean intersection over union (MIoU) of 91.15% and overall classification accuracy (OA) of 96.18% in complex scene recognition tasks in remote sensing images; furthermore, it also demonstrates excellent structural generalization ability in cross-domain transfer tests. This method has good scalability and theoretical transferability, providing an efficient and robust paradigm support for the intelligent interpretation of multi-source remote sensing images.