Research On Hyperspectral Images Classification Techniques For Vegetation Remote Sensing Monitoring

With the development and widespread use of remote sensing technologies,hyperspectral images are applied to the classification of ground objects and plant ecological monitoring.Currently,some classical methods of remote sensing data processing have been used to the hyperspectral image processing and analysis.However,these methods still cannot meet the requirements of the fine analysis for the classification of the change of ground objects and parametric inversion.Therefore,it needs to develop some image processing methods that different from the traditional ones,which can realize the classification of vegetation analysis and parameter inversion of hyperspectral images.The spectral curves of different ground objects can reflect the characteristics of different ground objects directly.Compared with other classification methods,the information of hyperspectral images can help to detect the changes of the ground objects accurately under the premise of choosing some appropriate analysis methods,which is of great significance to the fine analysis of remote sensing monitoring of vegetation types covered by ground objectsIn this paper,the analysis of vegetation ecological monitoring for hyperspectral images is achieved from three aspects,including the key spectral information extraction,migration feature fusion-based fine classification,and the change detection,and realize the classification and detection of changes in growth state of vegetation of hyperspectral images..Firstly,in view of the contradiction between the dimensionality reduction and the maximum utilization of spectral information in hyperspectral data,the feature selection algorithm based on local Fisher discriminant is studied.The algorithm considers the multimodal distribution structure of real hyperspectral data in the original feature space,selects a set of optimized feature subset,and keeps the local neighborhood structure of the samples in the same class.Based on the Fisher discriminant feature selection algorithm,this paper proposes a new classification method based on spectral key information fusion,which integrates the first derivative and the two order derivative of the spectrum with the original spectral features after determining the parameters of the spectral filtering order and the number of sampling intervals,and fuse the spatial features of different expressions,which helps to improve the classification accuracy of ground objects of hyperspectral images.Secondly,for the problem that the correct rate of classification is low,and the access and collection of image label is difficult when the labeled training data is insufficient.In this paper,a classification algorithm based on the feature fusion of knowledge transfer is introduced,which introduces the transfer learning into the network ensemble learning algorithm.The proposed algorithm can realize the classification of ground objects of hyperspectral images with fewer labels by utilizing the similar image label data which is different but related to the target data.The Ada Boost algorithm is adopted to construct a sample selection method for the source area to be migrated,and realizes the domain adaptive by using the sample migration.The experimental results show that,the classification method based on knowledge migration can still provide a high classification accuracy in the case of small sample.Finally,focuses on the problem that inversion of biochemical parameters by utilizing the spectral information of hyperspectral images,the inversion analysis of biochemical parameters is applied to the change detection of hyperspectral images.By analyzing the changes of biochemical parameters caused by vegetation changes,the quantitative inversion is performed for the representative chlorophyll parameters,water parameters and lignin parameters,and generates 9 component diagrams for each parameter's inversion result.By analyzing the components with the change information,the appropriate threshold is selected and the change area is obtained by dividing this component.Compared with the difference change detection method and the change detection method based on SVM classification,the proposed method can overcome the restrictions that the change detection result is limited by the classification results,high false detection rate,and the missed rate brought by the change detection,so as to realize the remote sensing monitoring and analysis of the growth state change of the same species vegetation of hyperspectral images.