An Improved Fusion in Extraction of Local and Global Features for Hyperspectral Image Classification

Hyperspectral imaging is now used in diverse applications like agriculture, forestry, and geology. Despite its versatility, accurately classifying hyperspectral data is challenging due to its high dimensionality and complexity. To address this problem Principal Component Analysis (PCA) might be used to extract the intrinsic features. However, PCA depends solely on the global variance of the input features. Only global variance may not be sufficient to achieve the informative features. This paper proposes an improved feature reduction approach that integrates t-distributed Stochastic Neighbor Embedding (t-SNE), PCA, and 3D Convolutional Neural Network (CNN) to capture the linear, nonlinear, local, and global characteristics in the generated output dataset.t-SNE is capable of capturing local data structures, while PCA grasps the global structures. Hence t-SNE reduced the dataset considering the local data structures of the original and then PCA is applied to reduce the input so that the output features are uncorrelated and contain most of the intrinsic structures. Therefore a fused dataset with both local and global features is obtained which leads to improved classification accuracy. Experimental outcomes on two real hyperspectral images demonstrate the outperformance of the proposed approach over conventional techniques. In the experiment, the proposed method obtained 99.77% accuracy which is the highest among the studied benchmark methods.