| Rice is one of the three major staple crops in the world and plays a crucial role in feeding a hungry world.As an important process of rice production,rice establishment methods have significant influence on production costs,yield and ecological environment.The traditional approaches to obtain rice establishment methods mainly rely on field visits,which are resource-intensive,time-consuming and unable to provide specific spatial distribution.At present,the rice establishment mapping methods based on satellite remote sensing have problems such as poor universality,scanty feature and unbalance between classification efficiency and accuracy.Accurate monitoring of rice establishment methods at large-scale still needs to be further explored.Therefore,the objectives of this study were to explore the differences in radar signals of paddy fields with different establishment methods and put forward a stable and universal method for mapping rice establishment methods at the county level.The performance of this method was further evaluated at the provincial level.This study will lay a foundation for fine classification of rice at large-scale.The main research contents and results of this thesis are as follows:(1)To develop object-oriented rice mapping methods;to analyze radar response signals of paddy rice fields with different establishment methods and propose mapping technology based on dual polarized imagery and machine learning method.Firstly,this study integrated high temporal-spatial resolution Planet imagery and multi-resolution segmentation algorithm to delineate field boundaries over six case cities/counties(Suining and Guanyun of Northern Jiangsu;Xinghua and Rugao of Central Jiangsu;Wujiang and Yixing of Southern Jiangsu)over Jiangsu province.Then,we further extracted the rice planting area by using the Sentinel-1 imagery during rice booting stage.At last,transplanted(TP)and direct-seeded(DS),two common rice establishment methods over Jiangsu province,were distinguished by combing the sensitive temporal and texture features of Sentinel-1imagery and Support Vector Machine(SVM).The results showed that the Sentinel-1 VH polarization imagery during rice booting stage could accurately extract the rice planting areas(OA=95%,Kappa=0.90)with the field boundary data.In addition,the relative errors between estimated and reported rice planting areas of case counties except for Xinghua are less than 10%.In terms of rice establishment methods classification,good performance was also achieved with OA of 86.29%and Kappa of 0.73.The distribution of rice planting methods obtained by our work is basically consistent with the truth.(2)To extract rice planting areas by integrating optical and SAR features at provincial level;to develop rice establishment mapping method applicable to large-scale based on cloud platform and clarify the contribution of radar texture features.In order to further evaluate the stability and transferability of the proposed method over large scales,this method was applied to map rice establishment methods over Jiangsu Province with Google Earth Engine(GEE)cloud platform.Firstly,Sentinel-1,the fusion of Sentinel-1 and Sentinel-2 images were input to a random forest classifier(RF)respectively with ground truth data for rice mapping over Jiangsu Province in 2020.Furthermore,depending on latter rice mapping results,we used the time series SAR imagery,texture features and SVM to distinguish TP from DS.We also evaluated the effect of texture features on the classification of rice planting methods by setting a contrast experiment.The results demonstrated that it is difficult to obtain accurate classification results merely based on mono-temporal SAR imagery(OA=75.11%,Kappa=0.66),and overestimation is obvious at provincial and county level.While further employing the optical vegetation indexes,the rice planting areas in Jiangsu Province could be extracted quickly and accurately(OA=95.01%,Kappa=0.93)and the estimated area exhibited stable fluctuations around reported data.At the county level,better correlation was obtained with the census reported data(R~2=0.90,RMSE=83.59km~2).Multi-temporal SAR features could separate TP from DS with poor performance(OA=80.26%,Kappa=0.60).When stacking the texture features,the OA increased to 82.96%and the PA of DS was improved by 7.6%.The misclassification from the edge of TP fields could be reduced effectively by integrating texture features.The proposed method that combing multi-temporal SAR imagery and texture features can accurately describe the distribution of rice planting methods over Jiangsu Province.It has great potential for rice establishment methods mapping over large scales.In conclusion,this study provide theoretical and technical support for all-weather and various scale rice establishment method mapping.Therefore,accurate information on rice management practices can help decision-makers to understand the current status of rice,optimize resource allocation and promote the sustainable development of rice production system.The proposed method has great potential in fine rice classification. |
PDF Full Text Request