Hyperspectral Image Classification Based on GS-CatBoost Model

Abstract

Abstract Aiming at the features of hyperspectral imges with large amount of redundant information of hyperspectral images, small correlation between features and categories, small sample size, and high-dimensions data, a model of hyperspectral image classification based on an improved mRMR band selection method and GS-CatBoost is proposed. First of all, using the mRMR improved by K-L to select the optimal band subset with high correlation between features and category, low redundancy, and large amount of information. Then, the Gabor filter is embedded in PCA and LDA combined dimensionality reduction to complete the extraction of space-spectrum joint features. Finally, using grid search to optimize the CatBoost algorithm to classify the extracted features. Experiments show that the proposed model fully and effectively extracts the space-spectrum joint features of hyperspectral images. In the Salinas Scene dataset, OA is 97.87%, AA is 99.01%, and Kappa coefficient is 0.9763.