Rice Growth Monitoring And Rice Yield Estimation In Poyang Lake Area Based On Polarization SAR

As an emerging technology,remote sensing has played an important role in large-scale monitoring of agricultural production.It can guide large-scale agricultural production and play a guiding role in international grain pricing by estimation of rice growth and rice yield through satellite image data.At present,the use of optical satellite images to monitor agricultural production has made great progress,however there was difficult to obtain optical data under cloudy and rainy weather conditions.Therefore,the all-weather observation capability of synthetic aperture radar is an effective supplement to optical data.The imaging principle and data processing method of synthetic aperture radar are quite different from optical data,and various monitoring and applications are still in the research stage.The purpose of this research was to deeply explore the radar wave scattering characteristics of rice in different phenological stages,reveal the SAR backscattering characteristics of the whole process of rice growth,and provide a theoretical basis for the inversion of rice biological parameters.This study provides technical,method and theoretical reference and basis for the growth status and yield estimation of crops in different periods by constructing rice biological parameters and crop yield estimation models,when there is a higher time,higher space and higher polarization in the future that characteristic SAR data is available.The goals of this research mainly include four aspects:first,to explore the optimal noise reduction method under different application scenarios of domestic GF-3 satellite SAR data,and to lay a solid foundation for the application of SAR data;second,to analyze the backscattering of different ground objects feature,improve rice classification and extraction methods,and mine the ground object classification capabilities of single-phase fullpolarization data and multi-temporal dual-polarization data;the third was to explore the inversion of rice biological parameters under different polarization data and their combinations.Methods;Fourth,construct a robust inversion model for rice yield estimation.In this study,subjective evaluation and objective evaluation indexes such as equivalent viewing number,edge preservation index and radiometric resolution were used to comprehensively evaluate the filtering results of SAR images.We carried out a classification study of SAR images for the QPSI full-polarization data of GF-3 and the dual-polarization multi-temporal data of Sentinel-1 respectively.The feature recognition and extraction capabilities of GF-3 data were deeply excavated.For the polarization characteristics of full polarization data,Pauli decomposition,Krogager decomposition and H/A/α decomposition models were used to analyze the polarization characteristics of SAR data.The feature matrix was decomposed and formed,and the feature-based optimization method was used to screen the fully polarized radar data to classify the optimal features.In this study,maximum likelihood method,support vector machine,neural network method and random forest method were used for classification comparison and accuracy verification.By using multiphase bipolar Sentinel-1 SAR data and using different phenological stages of rice growth data,to calculated the J-M distance to explore the optimal time-phase combination for early rice planting area extraction.We used unipolar linear model,polynomial model,exponential model and other single-polarization empirical models,dual-polarization empirical model,water cloud model and improved water cloud model and other biological models and mathematical statistical methods,analyzed with ground data collection,different rice scattering characteristics of the phenological period constructed the biological parameters and rice yield estimation models.The main findings of this paper includes:(1)Our study presents the optimal filtering method and processing window size for GF3 data in different study areas.It was recommended to use Kuan filtering in the strong texture area,and the processing window was 7×7,In the homogeneous area.It was recommended to use Kuan filter,Lee filter and homogeneous filter,among which Kuan filter 3×3,5×5 and 7×7 can achieve better results,but the processing window of Lee filter and mean filter should not be too large,which recommended 7×7 window size.It was recommended to use Lee filter and Frost filter in the edge area,and the window size was recommended to be 7×7,if the window was too small,it was easy to lose the linear target.(2)We determined the optimal time phase of rice extraction from multi-source polarized SAR data.Early rice extraction can be extracted by the combination of SAR data at threeleaf stage,direct seeding stage(corresponding to transplanting stage and returning to green stage),booting stage and milk stage,and the J-M distance between rice and lotus fields,woodlands,artificial surfaces,water bodies and wetlands can reach 1.832,1.979,1.890,1.975 and 1.999,respectively.(3)The classification results of different methods show that the results of random forest method was better than maximum likelihood method,support vector machine and neural network separation method.In this study,the temporal data determined by phase optimization were classified by random forest,and the overall interpretation accuracy was 94.3%,the Kappa coefficient was 0.93.(4)The extraction result of single-phase data was worse than that of time-phase data,but the fully polarized data still has high application potential.According to the feature vector group composed of three methods of decomposition of fully polarimetric data and four kinds of backscattering coefficients of the original image,the importance of features was evaluated,and the seven optimal features of fully polarimetric data classification in a single phase were optimized,namely:α,A,Red,H,Blue,Green and VV.The overall accuracy of random forest classification was 78.17%and Kappa coefficient was 0.738.(5)We constructed the optimal inversion model of rice microwave scattering.The inversion results show that the inversion result of VH polarization data was much higher than that of VV polarization data.The models with higher inversion accuracy included the empirical model of growth parameters using plant height,biomass and leaf area index as parameters,the water cloud model and the improved water cloud model.(6)We constructed the optimal inversion model of rice growth biological parameters.By using multi-temporal VH and VV dual polarization data for collaborative inversion,biomass and plant height PRD were 1.58 and 2.21;respectively,the improved water cloud model retrieved biomass and plant height PRD were 1.90 and 1.93.The improved water cloud model can explain the physical mechanism of rice microwave scattering and has greater advantages.(7)We constructed the inversion optimization model of rice yield.Rice biological parameters inverted by dual polarization empirical model and water cloud model can achieve ideal accuracy in rice yield inversion.Water cloud model can better explain rice microwave scattering mechanism and was more suitable for rice yield inversion.The final yield estimation result fitting determination coefficient was 0.7427,the%RMSE and PRD are 18.27%and 1.42 respectively,but on the whole,this method was feasible for regional rice yield estimation.