Semi Supervised Classification Of Hyperspectral Images Based On Multiple Classifiers

Hyperspectral remote sensing image,due to the characteristics of narrow band,can detect the spectral information of ground objects and express more abundant spectral curve features.Compared with multispectral images,hyperspectral images have more potential to distinguish ground objects.The band information hidden in the wide spectral range of multispectral images can be detected by hyperspectral sensors.Because of this,hyperspectral remote sensing images are widely used in agriculture,military,and geology.However,it is difficult to obtain ideal hyperspectral image classification results by current technology.Due to various complex situations of real objects,such as non-normal distribution of sample data,many noise points,large amount of data,different spectrum of the same object and same spectrum of foreign objects,it is difficult to classify real objects.These problems challenges to the increasing accuracy requirements of hyperspectral image interpretation.How to obtain universal,high-precision,and low-cost hyperspectral remote sensing image classification results is of great significance for the practical application of hyperspectral remote sensing images.The current research results show that it is difficult to obtain satisfactory classification results by using a single classification technology.How to build a complete hyperspectral image classification process with a small number of samples to achieve better classification results is a hot issue in mainstream research.This paper focuses on the classification of small sample hyperspectral images(1)In the case of small samples,a semi supervised hyperspectral image classification method is proposed by using filtering technology and multi classifier technology.The proposed method expands the training sample set by Multiple classifiers strategy,which provides more prior information,so that the subsequent classifiers learn more knowledge,and performs pre-processing and post-processing on hyperspectral images by filtering,which plays the role of de-noising and smoothing results respectively.The experimental results show that this method can improve the classification accuracy effectively compared with similar algorithms,and achieve good classification results.Experimental results on two datasets show the performance of the proposed algorithm.(2)In order to weaken the influence of noise points on classification results,this paper proposes a multi-scale super-pixel hyperspectral image fusion classification framework,which can extract multi-scale features of hyperspectral images,and effectively remove noise points through side window filtering,retain information of different scales,and finally segment and fuse multi-scale hyperspectral images through super-pixel technology.The experimental results on three datasets show that the framework is superior to the current cutting-edge classification algorithms.