Hyperspectral images (HSI) are intricate and
provide more spectral information than conventional images,
making them ideal for deep learning approaches that handle
3D data. In this paper, we present a classification method for
hyperspectral images (HSI) that begins with data
augmentation using flipping techniques during the
preprocessing phase. We then introduce a hybrid framework
that combines a Vision Transformer (ViT) with a 3D Region
Proposal Convolutional Neural Network (3D-RCNN). To
mitigate overfitting, especially when training data is scarce, we
incorporate a sequence aggregation layer. This methodology
captures more comprehensive spatial-spectral complementary
information while maintaining spectral integrity during
feature extraction. Ultimately, the k-Nearest Neighbors (KNN)
algorithm is utilized to classify the various classes within the
HSI images. Experimental results indicate that our method
significantly enhances classification accuracy across selected
datasets. |
