DenseBlock-Level Discriminative Multi-View Feature Fusion for Remote Sensing Scene Classification

Abstract

Remote sensing scene classification remains challenging because high-resolution aerial and satellite images often contain multi-scale objects, complex spatial layouts, high intra-class variability, and strong inter-class similarity. Although convolutional neural networks have achieved strong classification performance, most existing approaches primarily use the final-layer representation and therefore may not fully exploit the complementary texture, structural, and semantic information encoded at different network depths. This study presents a DenseBlock-level discriminative multi-view feature fusion framework for remote sensing scene classification. A DenseNet121 backbone is first fine-tuned on the target dataset, after which feature representations are extracted from its four DenseBlocks and treated as semantic-depth-specific views. The resulting descriptors are integrated using discriminative multiple canonical correlation analysis, which projects the heterogeneous views into class-aware latent subspaces by strengthening within-class relationships and reducing irrelevant cross-class correlations. The projected descriptors are subsequently concatenated and classified using a linear support vector machine. Experiments on the NWPU-RESISC45 and AID datasets show that the proposed framework consistently outperforms the fine-tuned DenseNet121 baseline, final-block features, direct concatenation, and PCA-based fusion across five random splits. The method improves the mean classification accuracy from 91.17% to 92.32% on NWPU-RESISC45 and from 92.29% to 93.76% on AID. These results indicate that supervised correlation-based integration of hierarchical DenseBlock representations provides a more discriminative and stable scene representation than conventional feature aggregation strategies