Polarimetric SAR Image Classification Based On Fine Description Of Spatio-temporal-scattering Information

Facing the application needs of land resource survey,agricultural activity planning,urban construction planning,etc.,image classification technology based on remote sensing platforms has become the main technique to achieve big data resource survey,agricultural modernization,and urban modernization.Polarimetric Synthetic Aperture Radar(Pol SAR),as an active imaging microwave remote sensing technology,has been successfully applied in both military and civil applications due to its all-weather and all-day earth observation capability as well as its penetration at limited heights.However,compared to optical sensors,the Pol SAR imaging mechanism is more complex and the image interpretation is more difficult.When using Pol SAR images for classification,it still faces a series of problems such as the confusion of the scattering mechanism,the influence of speckle noise,the large computational burden of massive data and the uncertainty of the scattering characteristics of crops,which restrict the practical application of Pol SAR images.To address these problems,this paper proposes a series of Pol SAR image classification algorithms based on the fine description of scattering information,temporal information,and spatial information,aiming to reduce the reliance of model training on labeled samples,increase the robustness and generalization of the model,and improve the accuracy of Pol SAR image classification.The details of researches in this paper are as follows.(1)Aiming at the confusion of the scattering mechanism of Pol SAR,two interpretable Pol SAR classification techniques are carried out in this paper.Firstly,an adaptive dimensional polarimetric feature decision tree classification algorithm is proposed,which applies the adaptive dimensional feature space to the decision tree nodes and uses the interpretable classification property of the decision tree to realize the fine description of land cover scattering mechanisms on the decision tree nodes by visualizing the classification process,and explores the mapping relationship between land cover scattering mechanisms and the polarimetric features.Secondly,a Pol SAR image classification algorithm based on hierarchical capsule network is proposed,and a uniform refinement description is designed for the scattering mechanisms,and the interpretable properties of the capsule network are exploited to uncover the intrinsic connection between deep features and scattering mechanisms.Finally,the effectiveness of proposed methods is confirmed by the classification results,and the accuracy of the description of scattering mechanisms is confirmed by transfer experiments.(2)To address the influence of coherent speckle noise and the large computational burden of massive data,the Pol SAR image classification technique based on non-local learning of spatial information is proposed.Specifically,firstly,we use superpixels instead of pixels for classification,which not only greatly reduces the number of computational units,but also enables the use of superpixels to mine fine pixel spatial contextual information,edge information,etc.Secondly,Wishart distance is extended to a similarity measure between superpixels and a superpixel-based graph representation is constructed to build the spatial topology of superpixels and capture the nonlocal spatial information of superpixels.Thirdly,the graph convolution network is used to convolve the superpixel-based graph representation and extract the deep semantic information of graph node superpixels by aggregating and transforming the information of graph nodes and their neighbors to achieve a rapid and fine classification framework for Pol SAR images.Finally,a multiscale superpixel fusion strategy is proposed for land covers with multiple scales,making full use of the spatial contextual information of large-scale superpixels and edge information of small-scale superpixels to improve the fine description of land cover scales without iterative training.(3)For crops with time-varying scattering mechanism,a two-element metric of time-varying information for time-series Pol SAR image classification technique is proposed.Specifically,firstly,a new tensor form is proposed to fine describe the multidimensional information of crops,including spatial dimension,polarimetric dimension and temporal dimension.Secondly,a two-element similarity metric of the tensor is defined,namely distance similarity and morphological similarity,and based on the two-element similarity metric,a tensor-based graph representation is constructed to establish the non-local spatial topology of the node tensor.Finally,the graph convolutional network is again used to achieve non-local learning of tensors and then fine classification of time-series Pol SAR images.Research content one proposes two interpretable classification algorithms,aiming to realize the fine description and verification of Pol SAR scattering information and provide feature basis for subsequent land cover classification.Research content two proposes a Pol SAR image classification based on nonlocal learning of spatial information,which mines the spatial context information of pixels and the spatial topological structure information of superpixels.And then using a graph convolution network to realize non-local learning,which breaks the limitations of local learning,solves the problem of pixel-by-pixel classification algorithm,and provides a new idea for the subsequent realization of Pol SAR image refinement classification.Based on the research results of research contents one and two,research content three introduces temporal dimensional change information,further broadens the dimension of fine description,defines a two-element similarity metric for timevarying information,and realizes the time-series Pol SAR image classification in combination with the classification algorithm of research content two.