Research On Hyperspectral Image Classification Model Based On CatBoost

With the development of image technology,hyperspectral images have received more and more attention in the field of remote sensing.At present,hyperspectral images have been widely used in fields such as agricultural development,military detectives,geological surveys,and weather forecasting.The classification of hyperspectral images is the basis of hyperspectral data processing.Because hyperspectral images are image data with high dimensions and large amounts of data,they contain a large amount of spatial information and spectral information.At the same time,there is a phenomenon of high correlation between data,which easily leads to a large amount of redundant data.It brings a huge challenge to the subsequent accurate classification.With the deepening of research on hyperspectral image classification,feature extraction and ensemble learning methods are widely used in hyperspectral image classification.Efficient feature extraction methods can extract the deep spatial and spectral features of hyperspectral images.Therefore,how to effectively extract and use these features for classification has become the focus of current research.This paper mainly uses band selection,combined dimensionality reduction,feature weighted fusion and CatBoost to study the classification of hyperspectral images.The specific research contents are as follows:(1)Aiming at the problems of high spectral dimension,strong band correlation,and small sample size,a GS-CatBoost hyperspectral image classification model based on band selection is proposed.Firstly,the m RMR algorithm improved by KL divergence is used as a new band selection method,and the optimal band subset is selected by the sequential search method,the original band information is retained to the greatest extent and the dimensionality reduction is achieved;Secondly,the principal components are extracted from the optimal band subset.On the one hand,PCA and LDA are used to reduce the dimensionality of the hyperspectral image to complete the extraction of the spectral features of the hyperspectral image,and at the same time,the intra-class spacing is reduced and the inter-class spacing is increased.On the other hand,the Gabor filter is used to achieve 5 scales and 8 directions spatial feature extraction;Finally,Catboost optimized by the grid search algorithm is used as a classifier to classify the space-spectrum joint features.The model is verified on the Salinas Scene dataset and the Indian Pines dataset.Through experiments,it can be found that the model has a better classification effect than the traditional model.(2)In order to further enrich the feature space and fully retain the effective features,on the basis of feature extraction and GS-CatBoost in(1),a GS-CatBoost hyperspectral classification model based on feature weighted fusion of genetic algorithm is proposed.First,perform band selection on the original data to extract spectral features.After band selection,perform Gabor spatial texture feature extraction,and perform PCA dimensionality reduction on the remaining bands to supplement spectral features;secondly,assign coefficient of variation weights to the three features,through improved The crossover and mutation of the genetic algorithm realize the screening and fusion of multiple features,and the CatBoost weighted classification is used as the result of the fitness function,and the optimal fusion features are continuously optimized.Finally,the fusion features are classified through the GS-CatBoost classification model.Experiments show that the proposed method makes full use of the multiple features of hyperspectral images,and the classification performance is greatly improved.