Effective Object Detection in Hyperspectral Images using Segmentation and Multibranch 2D-3D CNN Feature Fusion

In spite of the improvement of the Hyperspectral Image classification using a convolutional neural network, lesser training samples limit the classification performance. While 3D CNNs can extract combined spectral-spatial characteristics but need more complicated models, traditional 2-D CNN approaches underutilize the interband correlations in HSIs. Combining 3D-2D CNN multibranch feature fusion can take advantage of both the 3D and 2D CNN. Our proposed method implements Segmented Principal Component analysis (SPCA), minimum-Redundancy-Maximum-Relevance (mRMR) to exploit the local structure of the data and rank the features to form groups based on their significance. To extract the features as efficiently as possible, the shallow branch carries the least significant features, and the deep branch carries the top features. Our suggested model works better than the existing ones, attaining noteworthy accuracy of 99.12%, 99.94% and 100% on Indian Pines, University of Pavia and Salinas Scene datasets respectively.