Forest Classification Based On Compressed Sensing And Sparse Representation In Hyperspectral Remote Sensing Images

Hyperspectral remote sensing technology holds an important place in the military field not only,also be widely used in weather forecasting,territorial survey,agricultural output assessment,forest survey,geological prospecting,Marine forecasting,disaster monitoring,urban planning,mapping and so on various aspects,and with the global remote sensing satellite,communication satellite navigation and positioning satellite coordinates,It provides various information services for the national economic construction.In the application of forest remote sensing,hyperspectral remote sensing technology has also been widely used in the classification of forest species,monitoring and evaluation of forest diseases and insect pests,fire monitoring and extraction of forest resource change information by means of a large number of spectral information.Remote sensing technology provides a new scientific and effective management means for forest management and management.Although we can get the hyperspectral features by analyzing the reflectance spectrum of vegetation,the processing of hyperspectral data information is affected by the large amount of information and data redundancy in hyperspectral data.How to solve this problem has become the hotspot of hyperspectral remote sensing research and the development direction of remote sensing technology in the future.From the perspective of the development trend of forest forestry in China,in the whole process of data processing of hyperspectral remote sensing technology,the storage of hyperspectral images exists such problems as "Hughes state","isospectral enantiotopic" and"isomorphic isomerism".In addition,due to a large amount of data processing and the collection and transmission speed of hyperspectral data information and other problems,the middle and late image and image resolution caused a certain impact.After a detailed analysis and summary of the research status of hyperspectral images,this paper carried out a scientific study on efficient hyperspectral classification methods under the guidance of the basic theories of compressed sensing and sparse representation.The experiments of forest type classification and tree species classification on real hyperspectral remote sensing images in Wuyi Mountains show the effectiveness of the proposed method in classification.The specific research is as follows:Firstly,The hyperspectral remote sensing images in the study area were preprocessed,including data reading,striping,bad line repair,atmospheric correction,geometric precision correction,cropping and Mosaic,etc.The hyperspectral remote sensing data obtained by HJ-1A satellite consists of 115 bands,which are highly correlated with each other and have a lot of redundant information.Meanwhile,the number of bands has a certain impact on the speed and quality of data processing.Secondly,Aiming at the problem of hyperspectral remote sensing data redundancy,this paper uses the idea of block compressed sensing combined with TV method to establish a new hyperspectral remote sensing image reconstruction model based on GISMT compressed sensing and inter-spectral prediction.Firstly,the dimensionally reduced remote sensing image is divided into several non-overlapping equal-size image blocks based on the block sampling method.Then,according to the basic priori knowledge of remote sensing image,an improved joint sparse representation model is obtained to randomly sample each sub-image block.Finally,combined with the ALM reconstruction algorithm of TV method,the dimension-reduced hyperspectral remote sensing image is reconstructed from a small number of measured values obtained by compression sampling of all image blocks.Compared with the single ALM reconstruction algorithm,the time complexity of the proposed algorithm is smaller.The method of hyperspectral remote sensing image dimension reduction based on GISMT compressed sensing and inter-spectral prediction can be combined with forest classification.After the reconstruction,the sparsity of hyperspectral remote sensing image is more obvious,which lays a foundation for the subsequent classification algorithm based on compressed sensing and sparse representation.Thirdly,A forest classification method based on multi-feature dimension reduction and inter-spectral prediction reconstruction of hyperspectral remote sensing images was proposed.Firstly,the kernel function is used to map the samples to a high-dimensional feature space,and the feature selection is carried out according to the sparse representation coefficients in the high-dimensional feature space.In the process of classification,hyperspectral classification algorithm based on inter-spectral prediction reconstruction and multi-core support vector machine is adopted.Experimental results on standard hyperspectral data(sets)show that the proposed algorithm greatly improves the classification accuracy when learning from small samples.This algorithm can be used for general experimental data,and the classification effect is more than 95%,but for forest type classification only reaches more than 90%,which needs to be improved.The fourth,A window adaptive forest classification method for remote sensing images based on compressed sensing and sparse representation is proposed.This algorithm proposes a solution to the problem of the difficulty in high-precision classification caused by such phenomena as "foreign objects of the same spectrum,different objects of the same spectrum"and "dimensional disaster" in hyperspectral remote sensing images,as well as the problem of the failure to play the sparsity of reconstructed images.On the one hand,the shape adaptive search algorithm is studied to fully mine the image spectral structure information,and it is used to solve the coefficient optimization of the sparse representation model to improve the accuracy of representation.At the same time,the joint classification of similar pixels can effectively remove the classification results similar to "salt and pepper noise" generated by the isolated pixel classification,thus improving the overall classification performance.In the experiment,compared with the classical hyperspectral image classification methods,the experimental results show that the proposed method has a high classification accuracy in solving the classification problem of hyperspectral image forest categories,the average classification accuracy can reach more than 93%,and even some classification accuracy is more than 96%.The last,A forest classification method for hyperspectral remote sensing images based on unsupervised dictionary learning and sparse representation is proposed.This method is an unsupervised classification method.The dictionary learning method based on compressed sensing is used to replace the dictionary construction method based on the spectrum of training samples,which is helpful to construct a compact dictionary with stronger representation performance,thus improving the ability of accurate representation of complex spectral structures.The experimental results show that the proposed method has a great advantage in classification performance,and it can get a better classification effect when applied to Wuyiling tree species identification.Compared with other dictionary learning classification methods,this algorithm not only significantly improves the classification accuracy,reaching more than 95%,but also effectively reduces the running time of the algorithm.