| Soybean,as one of the major oil crops in the word,is among China’s largest imported agricultural products.Therefore,there is an urgent need to have an objective,detailed,and comprehensive understanding of the current situation of soybean production in China to contribute,and to formulate timely countermeasures to expand cultivation and promote the self-sufficiency rate.In view of the inadequacies of the previous studies on soybean identification by means of remote sensing,this study adopted GF-6 WFV image and multi-temporal Sentinel-2 multispectral as well as Sentinel-1 SAR images,and took the main soybean production areas,i.e.,Guoyang County in the northern plain of Anhui Province and Mingguang County in the Jianghuai region as the study areas,to explore the method of soybean identification under complex planting structure,diverse crop types and fragmented farmland landscape.The main tasks and achievements are as follows:(1)Based on the images obtained during the pod setting stage of soybean growing seasons in 2019 and 2020,this study selected typical soybean production areas in the northern Anhui plain and the Jianghuai hilly region,and adopted a stepwise hierarchical extraction strategy to obtain the spatial distribution of soybean planting areas.A set of filtering rules was first established and applied to remove non-cropland targets,and then the Relief F feature importance evaluation algorithm was employed combing with three machine-learning classifiers,namely Random Forest(RF),Support Vector Machine(SVM)and Back Propagation Neural Network(BPNN),to construct the combined Relief F-RF,Relief F-BPNN and Relief F-SVM models respectively to screen out the optimum feature-subsets for soybean identification,and the performance of the three models in soybean mapping was evaluated and the optimum model corresponding to each study area was then determined.In this study,eight ground samples were set up in Guoyang County and six in Mingguang County(5km × 5km in size),and the extraction results were evaluated based on the fine soybean distribution maps derived from Planet images.(2)Soybean identification in a typical main production area in northern Anhui Plain.In this study,a typical soybean production area,i.e.,Guoyang County,in northern Anhui Province,China,was selected as the study area,and single phase GF-6 WFV data and multi-temporal Sentinel-1/2 microwave and optical multispectral data were used to obtain the spatial distribution of soybean growing areas,respectively,through a stepwise hierarchical extraction strategy.First,time phase selection from multi-temporal data was conducted,the Sentinel-2 data acquired on 18 August,2019 9(early pod-setting stage)were judged to be more suitable for soybean extraction than other temporal phases according to the J-M distance.Based on the GF-6 WFV data and the specified Sentinel-2images,decision tree filtering rules were constructed respectively to remove the disturbance of non-crops such as water bodies,buildings,bare soil and trees,to obtain the general distribution of field vegetation.Furthermore,24 candidate features were generated from GF-6 WFV data,including eight spectral bands,nine vegetation indices,four texture features and three colour space components;while a total of 31 candidate features were obtained based on Sentinel-1/2 imagery:10 bands at 10 m and 20 m resolution,11 vegetation indices,4 texture features,and 6 microwave polarization features.With the support of ground survey samples of typical cover types,the Relief F feature weight evaluation algorithm was implemented coupled with specific machine learning classifier to obtain effective feature-subset for soybean identification,and soybean plots were extracted accordingly,then the 3-m resolution soybean distribution map obtained from Planet imagery was taken as the ground truth for validation.Results showed that for the GF-6 WFV data in Guoyang County,the Kappa coefficients generated from the best extraction model Relief F-SVM and optimum feature subset ranged from 0.69 to 0.80;the best extraction model Relief F-RF for Sentinel data outperformed the other two models,and the Kappa coefficients based on the optimum feature subset reached 0.77 to 0.85.(3)Soybean identification in a typical main production area of the Jianghuai hilly region.In this study,the same set of soybean planting areas extraction scheme was transferred to Mingguang in Jianghuai region,where the planting structure is more complex,to examine its application in this area.Based on multi-temporal Sentinel-2 data,the image acquired on 19 August,2020,early pod setting stage of soybean was determined to be more suitable for soybean extraction than other temporal phases by the J-M distance.Results showed that Relief F-SVM outperformed other two models for GF-6 WFV data,with Kappa ranging from 0.71 to 0.76;while the Kappa coefficients of the optimum extraction model Relief F-RF for Sentinel-1/2 data reached 0.72 to 0.87.(4)The extraction accuracy in Mingguang was generally lower than that in Guoyang County,mainly because the farmland here was more fragmented and d the geographical environment was more complex,which increased the difficulty of remote sensing mapping of soybean.Results from both study areas showed that Sentinel-2 red-edge bands,i.e.,B6 and B5,and NIR band(B8)were more sensitive to soybean,and the rededge B6 and NIR band(B4)of GF-6 WFV were also favorable to soybean identification;EVI,SAVI and CIgreen were more advantageous than other commonly used vegetation indices in terms of soybean mapping;and the texture feature “mean” was also effective.In addition,the extraction accuracy produced by the optimum feature subsets corresponding to the two data sources was significantly higher than the original spectral bands,comparable with that generated from the total feature,and the optimum model had obvious advantages in terms of data volume and computational overhead.Therefore,machine learning methods based on optimum features have great potential in remote sensing identification of soybean under complex planting structure.Figure [23] Table [21] Reference [115]... |
PDF Full Text Request