Research On Hyperspectral Image Classification Based On Rotation Forest Algorithm

Remote sensing images record rich spectral and spatial information of ground objects,which is the most intuitive representation of the overall appearance of ground objects.Using remote sensing images to identify and classify ground objects is of great significance in agricultural production,ground object detection,military reconnaissance and other fields.Machine learning methods have been widely used in remote sensing image classification.However,remote sensing images,such as hyperspectral images,have problems such as small number of samples,high feature dimensions and unbalanced categories,which lead to obvious limitations of traditional machine learning methods in the practical application of remote sensing image classification.Aiming at the problems encountered in remote sensing image classification,such as small samples and category imbalance,the following research works are carried out in this paper:Aiming at the problem of low classification accuracy of small and medium-sized sample data of hyperspectral images,the classification algorithm of small sample hyperspectral images based on space spectrum fusion and rotating forest was studied.The algorithm extracted the spatial features of hyperspectral images and fused the spatial features with the spectral features to obtain more detailed information of the images,and used the rotating forest method to classify the images.The experimental results show that the proposed algorithm has better classification performance in handling small sample hyperspectral image classification problems.The proposed algorithm is compared with the traditional classification algorithms such as K-neighborhood classification algorithm and support vector machine.Redundancy and noise characteristics brought by small sample data will have a negative impact on hyperspectral image classification.In order to increase the information content of small samples,this paper improves the hyperspectral image classification algorithm based on space spectrum fusion and rotating forest,and uses random forest to select features from the feature data set after space spectrum fusion,so as to reduce redundant features and avoid dimensional disaster.Then a small sample hyperspectral image classification algorithm based on random forest feature selection and rotating forest is implemented.The improved algorithm was applied to three groups of hyperspectral images,such as K-neighborhood learning algorithm of space spectrum fusion,space spectrum fusion and small sample hyperspectral images of rotating forest,to test the classification performance.The experimental results show that the random forest feature selection and the rotating forest algorithm can improve the classification accuracy of hyperspectral image in space spectrum fusion.Aiming at the problem of category imbalance existing in hyperspectral images,a synthetic minority oversampling based on the combination of empty spectrum is proposed technique(SMOTE)and the rotating forest imbalanced hyperspectral classification algorithm.SMOTE technology was adopted to process the category unbalanced data,and SMOTE algorithm was combined with rotating forest to balance the data set and increase the sample diversity.The classification performance tests were carried out on three groups of hyperspectral images,and the experimental results were compared with the classification algorithms such as random forest and support vector machine.The experimental results show that SMOTE and rotating forest unbalanced hyperspectral classification algorithm based on space spectrum combination has better classification performance when solving the problem of unbalanced hyperspectral image categories.