.png){kind=icon}
.png){kind=icon}
Return
Hyperspectral Image Classification Method Based on Prototype Feature-Driven Approach
.png){kind=icon}
PRE
AI
DOI：10.3788/LOP251951.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Objective Hyperspectral image (HSI) classification has become an important research direction in remote sensing, environmental monitoring, and precision agriculture due to its ability to capture fine-grained spectral information across hundreds of contiguous bands. However, two long-standing challenges restrict the effectiveness of existing methods: 1) limited annotated samples, which hinder deep learning models from fully exploiting spectral-spatial information, and 2) insufficient fusion of spatial and spectral features, which leads to redundancy and restricts discriminability. To address these issues, a novel hyperspectral classification model, termed prototype contrastive residual mixed convolutional network (PC-RMCN), is proposed. The model is designed to integrate prototype contrastive learning with residual mixed convolutional networks, thereby enhancing semantic guidance, improving spectral-spatial fusion, and ensuring better generalization under complex scenarios. Methods The proposed PC-RMCN consists of three main modules: 1) prototype extraction and feature enhancement module, which constructs class prototypes from limited labeled samples and employs contrastive learning to align features with their corresponding prototypes; 2) spatial-spectral feature extraction module, which uses multi-scale 3D convolutions to separately model spectral and spatial dimensions, and combines grouped convolution with pointwise convolution to achieve efficient fusion; and 3) residual attention enhancement module, which integrates both channel and spatial attention to reweight critical features and alleviate information dilution in deeper layers. In addition, a feedback mechanism is introduced between prototypes and convolutional layers: fused spectral-spatial features are fed back to dynamically update class prototypes, enabling a closed-loop interaction that improves prototype stability and enhances discriminability. The training process employs cross-entropy and contrastive losses jointly, optimized using backpropagation. Experiments are conducted on three widely used public datasets: Pavia University, Indian Pines, and Salinas. To comprehensively evaluate performance, overall accuracy (OA), average accuracy (AA), and Kappa coefficient are reported. Moreover, small-sample experiments and ablation studies are performed to validate robustness and the contribution of each module. Results and Discussions Experimental results demonstrate that PC-RMCN outperforms representative baselines including HybridSN, SSRN, DFSL, SSFTT, and DCLN across all three datasets. For instance, compared with HybridSN, PC-RMCN achieves improvements of 1.46 percentage points in OA, 4.77 percentage points in AA, and 2.68 percentage points in Kappa on the Pavia University dataset. Similarly, PC-RMCN surpasses SSRN, which attains a strong OA of 98.36% but suffers from redundancy in handling high-dimensional features, highlighting the advantage of prototype-guided fusion. Small-sample experiments further confirm the model's robustness. Even under a 2% training ratio, PC-RMCN consistently maintains superior performance compared to other methods, showing stronger generalization under data scarcity. Ablation experiments also reveal the contribution of individual components: removing prototype learning (Module 1), attention (Module 2), or simplifying spatial-spectral extraction (Module 3) each leads to noticeable accuracy drops, verifying the necessity of all modules. Furthermore, the closed-loop prototype feedback proves particularly beneficial, dynamically refining class prototypes during training and enabling the network to better capture discriminative spectral-spatial regions. The attention mechanism complements this by enhancing critical bands and spatial regions, mitigating redundancy, and ensuring more stable convergence. Finally, a discussion is carried out to highlight the broader implications. The results confirm the feasibility of PC-RMCN in addressing two key challenges in HSI classification: data scarcity and insufficient spectral-spatial fusion. Compared with existing methods (e. g., transformer-based models and self-supervised approaches), PC-RMCN achieves a better balance between computational efficiency and classification accuracy. Potential limitations include sensitivity of prototype initialization under extreme class imbalance, as well as stability of the feedback mechanism. Future research directions include integrating graph-based modeling, meta-learning strategies, and multimodal fusion to further enhance adaptability under cross-domain and complex distribution scenarios. Conclusions PC-RMCN provides an effective solution to the dual challenges of limited labeled data and inefficient spectral-spatial fusion in HSI classification. By combining prototype contrastive learning, residual mixed convolution, and attention mechanisms, the model achieves state-of-the-art performance on three benchmark datasets, with consistent improvements across OA, AA, and Kappa metrics. The proposed framework establishes a robust and efficient paradigm for hyperspectral classification, demonstrating strong generalization and stability.
Keywords:
image classification
hyperspectral image
prototype contrastive learning
residual convolutional network
attention mechanism
Journal
L
IF:
1
Papers:
596
Citations:
0