Research On Feature Selection Methods In The Field Of Cross-Scene Hyperspectral Remote Sensing Image Classification

Hyperspectral images usually contain a large number of spectral bands(features),and the redundant information in them can cause the Hughes phenomenon in the classification process,thereby reducing the classification accuracy of the image(Hughes phenomenon: the phenomenon where the performance of the classifier first increases and then decreases with the increase of band/feature dimensions).Therefore,reducing the redundancy of spectral information is a key step in the field of hyperspectral image classification.Feature extraction and feature selection are the two main methods to solve the above problems.Compared with feature extraction,feature selection methods retain most of the features of the original data without losing valuable details.However,existing feature selection methods based on a single scene perform poorly in certain scenarios(domains)with insufficient labeled samples.Therefore,it is still very challenging to adopt efficient feature selection methods to select the optimal feature subset of the source and target scenes,and use the sample information of the source scene to assist in the classification of the target scene,in order to improve the classification accuracy of the target scene image as much as possible.To address the above issues,this paper proposes a new cross scene feature selection algorithm for hyperspectral images: the cross scene feature selection algorithm based on improved ant colony optimization algorithm(IMACA-CSFS).The main research of this paper is as follows:(1)A priority sorting-based ant colony strategy is proposed,which enables subsequent search processes to focus on the global optimal solution(optimal feature subset)found in the previous iteration for iterative search.Compared with other cross scene feature selection methods,this strategy can obtain more accurate optimal feature subsets of the source and target scenes;An ant colony strategy based on elite ants is proposed,which improves the defect that the optimal solution of the original ant colony algorithm based on single scene feature selection converges too slowly in each iteration process,further speeds up the convergence speed of the global optimal solution,obtains the optimal feature subset of the source scene and target scene with higher efficiency for training classifiers,and further improves the update strategy of pheromone in the ant colony optimization algorithm,and experiments were conducted on three publicly available hyperspectral image data pairs,Huston 2018-Huston 2013,Pavia U-Pavia C,and Shanghai-Hangzhou,verifying the feasibility and superiority of the two ant colony strategies.(2)A target function model based on both source and target scenes was designed,which simultaneously considers the distinguishability of different types of ground objects in the target scene and the consistency of selected features between the two scenes.The feature selection method IMACA-FS based on a single scene was successfully applied to the field of cross scene hyperspectral image feature selection and image classification,effectively weakening the impact of spectral drift on cross scene image classification,This greatly improves the feature selection and image classification accuracy of the target scene.The experiments conducted on three publicly available cross scene hyperspectral data pairs have demonstrated that the new method IMACA-CSFS proposed in this paper has better performance in cross scene feature selection and cross scene hyperspectral image classification.In conclusion,the IMACA-CSFS method proposed in this paper can further improve the updating strategy of pheromone in the original ant colony algorithm,inherit the original characteristics and advantages of IMACA-FS method,effectively weaken the impact of spectral offset on cross scene hyperspectral image classification,and greatly improve the feature selection accuracy and image classification accuracy of the target scene.