Object-Level Crop Classification Based On The Time-Varying Characteristics With Time-Series Polarimetric SAR Data

Facing the national food security and economic development,regional crop classification with remote sensing platform has become the main technique to achieve high-quality agricultural development and modernization.Polarimetric synthetic aperture radar(Pol SAR)is a rapidly developing active microwave remote sensing technology.It can realize the all-time and all-weather continuous observation,obtain multi polarization observation by transmitting and receiving variable polarization electromagnetic signals,and characterize the morphological structure,arrangement and distribution,dielectric constant characteristics of crops.Based on this advantage,Pol SAR has been proved to be able to effectively characterize the growth state of crops and distinguish different crop types.Furthermore,the vigorous development of spaceborne Pol SAR platform provides an opportunity to improve the accuracy of crop classification and parameter inversion,making time-series polarimetric SAR(time-series Pol SAR)technology concerned by more and more countries and institutions.By using time-series Pol SAR images,mining the time-varying characteristics of polarimetric information,coupling the growth law and phenological evolution of crops,facing the main problems in agricultural application,it has become the mainstream research in remote sensing to build a universal and generalized intelligent crop classification method.Considering the speckle noise in Pol SAR images and the calculation burden of time-series massive data,the object-level time-series Pol SAR classification method has been studied by many scholars.It not only improves the signal-to-noise ratio(SNR)of Pol SAR data based on the average of homogeneous pixels,but also greatly reduces the number of data processing units,which effectively improves classification accuracy and efficiency.As for agricultural applications,crops are mostly distributed in parcels,and the object-level classification is suitable for practical requirements.High-precision crop classification results depend on accurate object generation and robust classification.The former focuses on the authenticity of farmland geometric locations,mainly involving edge detection and object segmentation methods;The latter focuses on the accuracy of crop type discrimination,closely related to feature extraction and classification methods.However,there are many problems in timeseries Pol SAR agricultural applications,including high false alarm rate of edge detection results,poor boundary attachment of segmented objects,low discrimination ability of time-varying features to crop types and weak generalization of classifier.In view of this,this paper takes time-varying Pol SAR characteristic as the starting point,tackles the difficult problems in time-series polarimetric similarity measurement and crop classification with time-series alignment algorithms,studies four parts,including the edge detection method of farmland parcel based on time-varying polarimetric characteristics,the hierarchical object segmentation method based on spatial-temporal context information and edge constraints,the high-quality time-varying feature extraction and the crop classification method based on the time-series alignment and machine learning method.Research part 1 and part 2 aim to accurately detect the geometric location of farmland parcels,and research part 3 and part 4 aim to accurately identify the crop type.Four parts are progressive and closely related.The crop classification is realized on the basis of the segmented objects generated by the part 1 and part 2,aiming to break through the bottleneck of crop classification caused by complex time-varying characteristics and phenological uncertainty,construct a complete system of object-level timeseries Pol SAR classification to satisfy the rapid and timely crop monitoring and classification.The main contributions and innovations of this paper are as follows:(1)A time-series Pol SAR edge detection method of farmland parcel based on time-varying characteristics is proposed.The method improves the discrimination of polarimetric similarity to objects with different temporal evolutions,increases the SNR of the edge strength and improves the edge detection accuracy.First,aiming at the problem that the traditional polarimetric similarity is difficult to identify the edge between the objects with different temporal evolutions,we propose a timeseries polarimetric similarity with root mean square,which improves the sensitivity of similarity to objects with different temporal evolutions and the edge recognition rate.Then,aiming at the problems of low contrast of edge intensity and high false alarm rate in traditional methods,a joint edge intensity is proposed with the weight of spatial-temporal homogeneity factor,to enhance the edge strength of pixels within edge-region,reduce the background noise,and effectively improve the edge detection accuracy.By using the time-series Radarsat-2 quad-polarization SAR data in Flevoland and the time-series Sentinel-1 dual-polarization SAR data in Jinchang,it has been demonstrated that,compared with most single-phase results,the improvement of boundary recall of the proposed method is over8% and 15% in the full-polarization dataset and the dual-polarization dataset,respectively,and the improvement of F1 score is over 7% and 5%,respectively.The proposed time-series polarimetric similarity can detect more parcel boundaries than traditional method,and the signal-to-noise ratio of the proposed joint edge strength is also better than traditional methods.(2)A hierarchical object segmentation method based on spatialtemporal context similarity and edge constraints is proposed.It improves the accuracy of time-series Pol SAR segmentation map model with the spatial-temporal neighboring information,and establishes a robust minimum heterogeneous region merging criterion under edge and shape constraints,and improves the accuracy of multi-scale object segmentation.First,under the framework of time-series Pol SAR similarity measurement based on root mean square,the spatial-temporal context similarity is constructed by temporal and spatial neighborhood information.The edge weight between the nodes in map model is calculated by the local spatial-temporal blocks,which reduces the noise and improves the accuracy of the map model in complex scenes and reduces under-segmentation and over-segmentation errors.Then,based on the minimum spanning tree domain generated by Kruskal algorithm,a minimum heterogeneous region merging criterion for time-series Pol SAR data is proposed by considering edge and shape constraints,which is referred by the polarimetric likelihood statistical theory.The novel criterion can preferentially merge the homogeneous pixels within one cropland parcel.Validated by the time-series Radarsat-2 quad-polarization SAR data in Flevoland and the time-series Sentinel-1 dual-polarization SAR data in Jinchang,the segmentation results of different scales generated by our method can better adhere to the edge information of the parcel,and the proposed spatial-temporal neighborhood information and edge constraints can effectively improve object segmentation accuracy.(3)A multi-feature weighted joint time-series Pol SAR crop classification method is proposed based on dual-similarity with timeseries alignment.The shape similarity based on the time-series alignment is extended to the dual similarity.Then,the weighted joint multi-feature similarity is constructed based on the class separation degree with dual-similarities,to effectively improve the separation ability of time-varying features to different crop types.Caused by the random variations of agroclimatic conditions and agricultural practices,two neighboring parcels of the same type may have different phenological evolutions,which generates the classification error.In order to suppress the negative impact of phenological uncertainty on regional crop classification,the dynamic time warping(DTW)alignment is introduced to improve crop classification.Two improvements are proposed to improve the class separation ability of time-varying features,where traditional methods only consider shape similarity and single feature type.In this paper,the scattering feature difference of microwave signal in crops is introduced,the feature similarity is proposed,and the dual-similarity is constructed together with shape similarity.Then,the class separation degree based on the dual-similarity is constructed,which is used by weights to construct the joint similarity of multiple features.The joint similarity can avoid the effects of adverse features and improve the discrimination ability of timeseries Pol SAR features.Verified by the time-series Sentinel-1 images acquired in 2018 and 2019 in Jinchang,our method can obtain the maximum improvements of 31% and 25% in 2018 and 2019,respectively,compared with the classical nearest neighbor classification method with DTW alignment and single feature type.Furthermore,compared with the traditional shape similarity,the overall classification accuracy of the proposed dual-similarity in 2018 and 2019 was improved by 2% and 5%,respectively.(4)A crop classification method based on pairwise proximity function support vector machine(ppf SVM)classifier with time-series alignment kernel is proposed.The time-weighted shape DTW(TWshape DTW)alignment is proposed to reduce the matching error,and the ppf SVM classifier can improve the classification generalization,which can effectively improve the classification accuracy.First,for the disadvantage of the DTW alignment that the local shape of the curves and temporal ranges are neglected,the TWshape DTW is proposed to reduce the matching errors and improves the intra-class similarity of samples owing to the same type.Then,the ppf SVM classifier based on time-series alignment kernel is proposed,which constructs a positive definite time-series alignment kernel matrix with dual-similarity of multiple features.The classification accuracy can be effectively improved with the advantages of SVM classifier taking into account structural risk and empirical risk,simultaneously.With the time-series Sentinel-1 data in the study area of Jinchang,our overall accuracies in 2018 and 2019 exceed 90%,and the improvement of overall accuracy in 2018 and 2019 obtain about 6% and 12%,respectively,compared with the proposed method with DTW alignment proposed in the part 5.This proves the effectiveness of the TWshape DTW alignment and the ppf SVM classifier.Furthermore,this improvement is more effective in small samples than enough samples.