Rice Information Extraction Studies Based On Multi-source Remote Sensing Data Integrating And Data Assimilating

Rice is one of the most important crops in China,which has the largest planting area and the highest yield.The planting acreage,overall output and per unit area yield occupy primacy in all foodstuff.Extracting the information of rice production accurately and grasping the crop production status have an important realistic significance for food security,social stability,rural employment and increasing farmers’ income.Rice information extraction is a synthetic application of remote sensing（RS）,geographic information system（GIS）and global positioning system（GPS）.In this study,we solved some specific questions in rice information extraction in relatively fragmented area in China based on multi-source and multi-temporal middle to high spatial resolution remote sensing data and extensive field campaign data using data integrating and data assimilating methods.Quality assessment and some quantitative researches have been made on influnrence factors during the process in this dissertation to some extent.The main research contents and achievements of this dissertation are as follows:（1）Field crop canopy spectroscopic data and field observation data were intergrated to explore the growth rhythm of single-cropped rice（SCR）.The discriminant analysis using Wilks’s lambda coefficient showed that the two-band enhanced vegetation index（EVI2）can reflecte the difference between rice field and other land cover types better.By analyzed the Savitzky-Golay（S-G）filtered HJ-1 CCD EVI2 time-series data of five main land cover types（rice,trees,water bodies,economic crops and other nonvegetated areas）in study area,a stepwise classification strategy utilizing the EVI2 signatures during key phenology stages,i.e.,the transplanting and the vegetative to reproductive transition phases,of the SCR was proposed.The overall classification accuracy and Kappa coefficient for five field sites were 91.68%and 0.79,respectively.Compared with the statistical data of the local agriculture department,the relative classification accuracy was about 91.2%.The overall classification accuracy was 6.80%and 10.89%higher than the results of one typical parametric classification algorithm（the Maximum Likelihood Classifier,MLC）and one typical nonparametric classification algorithm（Support Vector Machines,SVM）,respectively.Due to the fragmented land use composition in the study area,we also assessed the influence of mixed-pixel and boundary effects quantified by using the landscape indices and it showed that the classification accuracy ratio of rice field was positively correlated with its compactness.The result showed that by making full use of the key phenological information,and under the support of high-temporal resolution remote sensing data,e.g.,HJ-1A/B,the SCR can be mapped at a relative high confidence.As large part of the classification error can be attributed to the influence of mixed-pixels where the area proportion of rice field was less than 75%,and most of the mixed-pixels concentrated at the boundaries of the rice fields.（2）By integrating the vegetation index（Ⅵ）of HJ-I CCD and Landsat-8 OLI data via the ordinary least-squares method（OLS）,the key phenology parameters（transplanting/heading/maturity stages）of SCR in study area can be estimated more accurately than a single data source.The response ability of vegetation index to target objects differs,compared with the normalized difference vegetation index（NDVI）,EVI2 was more stable and comparable between the two sensors.Compared with the field observed phenological data of the SCR,the integratedⅥtime-series had a relatively low RMSE,which EVI2 is better than NDVI.In the process of integrating different sensors for higher temporal resolution time-series vegetation index,the inconsistency between the two sensors could be attributed to systematic and unsystematic differences quantitatively described by agreement analysis.The systematic difference between the two sensors could be minimized by OLS,while the unsystematic difference could be effectively attenuated using 5×5 filter window,which could minimize the influence of the environmental factors,thereby reducing the uncertainties generated from data integrating process.（3）Based on field campaign data in 2012-2013 as modeling and validation data, a near real-time dynamic mapping method for SCR growth parameters at a regional scale was proposed via multi-temporal remote sensing images.By cloud removing process and max value composite method,the 10-day HJ-1 CCD image composites,giving a total of three image composites per month were generated for growth parameters retrieval.The field measurements included leaf area index（LAI）,above ground biomass（AGB）and plant density,were strictly controlled during measurement.Two machine learning methods,i.e.back propagation neural network（BPNN）and support vector machine（SVM）were applied in LAI model construction.The LAI empirical equations were established by dividing the whole SCR growth period into before-heading and after-heading stages.EVI2-BPNN regression was selected for the before-heading LAI estimation,and the NDVI-SVM was selected for the after-heading-LAI estimation.That means EVI2 performed better at the fast growth stage（before-heading）whereas NDVI showed better accuracy at the after-heading stage（relatively slower growth）.Cumulative VIs were proved suitable for the estimation of AGB.Cumulative NDVI based on the quadratic polynomial fit function was adopted for the all-stage AGB prediction,with the coefficient of determination values is 0.93.The results demonstrated the potential of using multi-temporal images in rice growth monitoring.Machine learning methods provided a useful exploratory tool for improvement on the relationships between different combinations of reflectance and crop variables.By selecting the appropriate vegetation index and the models,the accuracy of rice growth parameters can be improved and provide valuable visual information services for local plantation management.（4）The WOFOST crop model parameters localization was conducted by combining field campaign data and FSEOPT in the study area.The localized model parameter showed that the crop model represents a reasonable rice growth trend and estimates the potential yield at point scale.A shorter time interval of assimilation state variable led to a better result when assimilating time-series LAI and WOFOST crop model by using Ensemble Kalman Filter（EnKF）methods.The 10-day LAI was selected for regional rice yield estimation by taking balancing the assimilation efficiency and precision into consideration.The yield distribution information in the study area was estimated by using rice area map and rice phenology map as model input parameters,10-day remote sensing retrieved LAI as assimilation data.The simulated model results and measured data had a good consistency.Compared with the measured yield,R2 and RMSE were 0.66 and 1.61 ton·ha^-1,respectively.The research result showed that,on the premise of improving the precision of each step in rice information extraction process,WOFOST model can be used in regional scale rice yield potential estimation and provide valuable information for the cultivation management improvement and grain production special zone planning.