Hyperspectral image provides hundreds of continuous spectral measurements, and selecting a subset of bands with distinct and independent features from these numerous channels is a crucial problem. In recent years, although scholars have proposed many methods for band selection, most of these methods only focus on the information content of the bands or the redundancy between the selected bands. To comprehensively consider the redundancy and information entropy between bands, we propose a hyperspectral band selection method that considers band-sharing neighbors. This method consists of two parts: subspace partitioning and weight ranking. By sharing neighbors between bands, we optimize the spatial pre-segmentation points to appropriate positions, to maximize the differences between different subspaces. In the band selection stage, we comprehensively consider factors such as local density, information entropy, and signal-to-noise ratio to select the optimal band subset. Through extensive comparative experiments on three public datasets, we demonstrate a significant improvement in accuracy and efficiency with this method.

This paper presents a shared nearest neighbor band selection method based on local density, enabling rapid band selection for hyperspectral images while maintaining accuracy. Specifically, the proposed method comprises two steps: subspace partitioning and comprehensive weighted ranking. During subspace partitioning, the method first pre-partitions the hyperspectral image bands into subspaces by evenly dividing them. It then dynamically adjusts the interval points between subspaces by considering the correlation of shared nearest neighbors among the central bands of each subspace. After completing the subspace partitioning, the method comprehensively considers local density, information entropy, and signal-to-noise ratio to select the optimal subset of bands. Compared to other band selection methods, the approach proposed in this paper has two main advantages. First, subspace partitioning does not require multiple iterations over the redundancy and correlation of each band, significantly reducing computation time. Second, during the weighted ranking process, multiple influencing factors are comprehensively considered, thereby avoiding confusion in information entropy calculations caused by atmospheric noise.

The method proposed in this paper was extensively compared with common band selection methods on three public datasets using support vector machine (SVM) and K-nearest neighbor (KNN) classifiers. The results demonstrate that the applicability and accuracy of our method. Through experiments, the optimal parameter combinations for our method on different datasets were determined. The classification accuracy of our method with different parameters using SVM and KNN classifiers is shown in Tables 1, 2, and 3. In ablation experiments, the structure of our proposed method was replaced with that of other competitive methods for comparison. The results, shown in Fig. 8, indicate that replacing the clustering method and ranking strategy led to a decrease in classification accuracy for both SVM and KNN classifiers, with the clustering method having a more significant impact. Specifically, replacing the clustering method with PIENL (Pearson correlation coefficient, information entropy and noise level) resulted in a decrease in overall accuracy (OA) values by an average of 1% to 4%, with the KNN classifier on the Pavia University Scene dataset showing the largest variation by up to 4.2%. As for the ranking strategy, the modified method also showed a decrease in accuracy, but the average decrease remained within 1%. The performance of each method was evaluated by comparing OA, average overall accuracy (AOA), and runtime. As shown in Fig. 9 and Table 5, our proposed method can quickly and accurately identify hyperspectral band subsets with more information content and lower redundancy.

This paper proposes an efficient and accurate solution for the band selection problem in hyperspectral images based on shared nearest neighbors between bands. The main contributions of this paper are as follows: constructing a correlation matrix using Euclidean distance and grouping bands based on shared nearest neighbors, thereby dividing the bands into multiple reasonable groups. Maximizing inter-group differences and intra-group similarity by considering local density, thus optimizing the partition points between different groups. During weighted ranking, comprehensively considering image information entropy and signal-to-noise ratio to precisely select a subset of bands from within the groups that have high information content, low redundancy, and high signal-to-noise ratio. Extensive experiments were conducted on three public hyperspectral image datasets using two classifiers, and the results demonstrate the robustness and effectiveness of the proposed method. For future work, we plan to further optimize this method in two aspects: automatically evaluating the size of the selected band subset using specific methods to avoid information loss or redundancy. Continuing to optimize the algorithm to accelerate its runtime.