Relative Water Content Retrieval And Refined Classification Of Ground Objects In Hyperspectral Image

Hyperspectral remote sensing technology of earth observation provides data support for the quantitative and refined interpretation of ground objects.In addition to spatial information,such as the target’s geometric texture,hyperspectral images also contain spectral information that can be used to detect changes in inherent attributes like biophysical and biochemical(namely physicochemical)properties.The extraction of physicochemical diagnostic spectral signatures enables the relative relationship analysis and quantitative retrieval of typical physicochemical parameters,such as the water content of objects,and also enhances the intraclass separability of ground objects with similar spatial-spectral features.By introducing spectral diagnostic signature into the classification of spatial-spectral statistical features,it is capable of further improving the intraclass accuracy of ground objects with similar spatial-spectral features while improving the interclass accuracy,which also achieves the refined classification of intraclass and interclass mixed hyperspectral images.In relative water content retrieval and refined classification of hyperspectral tasks,low band information utilization and weak expression of single diagnostic spectral signature have an impact on the improvement of retrieval accuracy,and the lack of labeled data and the differential distribution of data in other sources limit the optimization of spatial-spectral feature extraction in transfer learning classification.Moreover,the structure limitation of deep feature extraction makes it challenging to retain and retrieve physicochemical diagnostic spectral signatures for refined classification.Thus,this thesis focuses on the imagespectral characteristics of hyperspectral data and conducts related research,examining the expression of spectral diagnostic signatures and spatial-spectral statistical features in physicochemical parameter retrieval and classification,respectively,and carrying out research on relative water content diagnostic signature extraction and parameter retrieval,as well as the transfer optimization of spatial-spectral feature extraction in classification.By the introduction of relative water content signatures into the spatial-spectral features based classification,the research on refined classification of hyperspectral images has been achieved.The main research contents of this thesis are as follows:In order to analyze the extraction and expression of spectral diagnostic signatures and spatial-spectral statistical features in hyperspectral image,this thesis researches the process of physicochemical parameter retrieval and spatial-spectral classification.By analyzing the correlation between spectral curves and physicochemical properties in typical vegetation,the spectral diagnostic signatures of different physicochemical parameters are abstracted and summarized,and the associated parameter retrieval models are established.The effectiveness of the summarized step-type and morphological signatures for parameter retrieval is verified by experiments.By analyzing the process of extracting spatial-spectral features from hyperspectral images,a dimensionality-varied convolutional neural network module for fused spatial-spectral feature extraction is established to improve the accuracy and efficiency of classification,which is also verified by experiments.The aforementioned research provides a theoretical and model basis for the following implementation of diagnostic signature extraction in relative water content retrieval and spatial-spectral feature transfer optimization in classification.In order to improve the weak expression of single diagnostic signature and insufficient utilization of band information in relative water content retrieval,this thesis researches the diagnostic signature extraction and retrieval algorithm of hyperspectral relative water content,and proposes a convolutional neural network retrieval algorithm constrained by prior information of red edge slope.By extracting and fusing the watersensitive red edge band signatures and the learning signatures of convolutional neural networks,the utilization of spectral bands and the diversity of signatures are improved,and the diagnostic expression ability of signatures for water content is enhanced.The accuracy of retrieval is optimized,and quantitative mapping is achieved.The effectiveness of this algorithm is also demonstrated by experiments.The aforementioned research provides accurate diagnostic signatures of relative water content for following refined classification to improve the recognition ability of intraclass ground objects.In order to improve the effectiveness of spatial-spectral statistical feature extraction on different distribution data,this thesis researches the optimization of feature extraction capabilities of hyperspectral image classification,and proposes a transfer optimization algorithm for spatial-spectral feature extraction based on homogeneous and heterogeneous data fusion in the source domain.By establishing mapping relationships from multiple homogeneous and heterogeneous hyperspectral data in the source to the target domains,the capacity of spatial-spectral feature extraction in classification models is enhanced for different data.The correlation and difference between the homogeneous and heterogeneous data are used to enhance the category sensitivity and distribution robustness of the model on target data,which improves the interclass classification accuracy and optimizes the spatial-spectral feature extraction ability.This algorithm has been validated by experiments,and provides abundant high-precision spatial-spectral statistical features for following refined classification to improve the interclass classification ability.In order to improve the utilization of spectral diagnostic signatures and retain the weak features in refined classification models,this thesis researches the joint extraction of diagnostic signatures and spatial-spectral features,and proposes a refined classification model based on the relative water content retrieval and feature transfer optimization.By establishing a basic classification framework on multiscale and multiattention feature extraction,and utilizing the research results of relative water content retrieval and transfer optimization of spatial-spectral feature extraction,the fusion expression of spectral diagnostic signatures and spatial-spectral statistical features is realized,which improves the interclass accuracy and further improve the intraclass accuracy.The refined classification of hyperspectral images is achieved by this algorithm.The effectiveness of the refined classification model has been proven by experiments,and the outstanding performance of proposed algorithm in hyperspectral image refined classification tasks has also been demonstrated.In summary,the focus of this thesis is on the analysis and extraction of hyperspectral image-spectral characteristics to improve the diagnostic signature expression of relative water content,as well as the spatial-spectral statistical feature extraction by transfer optimization,which achieves the refined classification of hyperspectral image.The research of this thesis provides important support for the quantitative and refined interpretation of hyperspectral images.