Hyperspectral Image: A Fusion-Based Extraction and Classification Utilizing Centernet Technique

Due to the large dimensionality and few training samples with labels, classifying hyperspectral images (HSIs) is a complex process. We proposed an innovative approach to HSI categorization based on a multi-view deep learning approach that integrates spectral and spatial characteristics by utilizing a few labeled instances to overcome these difficulties. To extract a set of spectral and spatial information, we initially process the original HSI. Every spectral vector is the spectrum characteristic of a single picture pixel. A deep Siamese convolutional neural network is employed to extract the spatial features from the data. This network aims to minimize the high dimensionality of the data by accounting for each pixel’s immediate neighborhood. The spectral local texture characteristics are extracted employing the sparse autoencoder, which has a minimal computational burden to create multi-feature vectors. Next, we propose a multi-view approach. The non-subsampled contourlet transform model can be used to combine the spectral and spatial information extracted from the HSI into a single latent representation space. The categories are categorized using a deep learning-based system called CenterNet. Using the datasets from Salinas, WHU-OHS, and Pavia University, the effectiveness of the suggested method is verified. When compared to current state-of-the-art techniques, the comparative experiment results show that the suggested strategy executes better in terms of generalization for small samples and may produce higher categorization accuracy. It presents a novel viewpoint on the processing of HSIs.