Research On Extraction Method Of Soybean Planting Area Based On Multi-source Remote Sensing Data

Soybean is the main grain and oil crop of China,obtaining the spatial distribution of soybean planting area timely and accurately is important for yield estimation,crop-damage warning and agricultural policy adjustments.At present,remote-sensing technology is an effective approach to obtain crop planting area which can conduct a real-time dynamic monitoring of farmland in an objective,efficient and low-cost manner.Firstly,this thesis selected two towns(Qingtuan and Longshan)situated in typical soybean producing areas in North Anhui plain as the satellite remote-sensing study area,and multi-temporal Sentinel-2images were employed to obtain the spatial distribution of soybean in the 2019 growing season.The extraction effect of Sentinel-2 images was evaluated by unmanned aerial vehicle(UAV)images covering six ground samples deployed in the study area.Then,from the satellite level to the UAV level,the RGB image of UAV(1km×1km)located in Yimen Town was used to carry out the extraction of soybean planting area in a small area.The main work of this thesis is as follows:(1)Study on extraction of soybean planting area based on Sentinel-2 data.Based on the three images,the image obtained on August 18(early pod-setting stage of soybean)was determined to be the most applicable to soybean mapping by means of the Jeffries–Matusita(JM)distance.Subsequently,three machine-learning algorithms,i.e.,random forest(RF),support vector machine(SVM)and back-propagation neural network(BPNN)were employed and their respective performance in soybean mapping was evaluated by 254 ground truth plots.It appeared that RF with a Kappa of 0.83 was superior to the other two methods.Furthermore,20 candidate features containing the reflectance of 10 spectral bands(spatial resolution is less than or equal to 20 m)and 10 remote-sensing indices were input into RF algorithm to conduct an important assessment.Seven features were screened out and served as the optimum subset,the optimum feature-subset could achieve a high-accuracy result,with a reduction of data volume by 65% compared with the total 20 features,which also overrode the performance of10 spectral bands.Therefore,feature-optimization had great potential in the identification of soybean.(2)Study on extraction of soybean planting area based on hierarchical extraction strategy.In order to further improve the effect of soybean extraction,a set of decision tree filtering rules were established firstly based on the Sentinel-2 image at the optimum time phase to eliminate non-agricultural cover types and thus obtain the overall distribution of field vegetation.Then,19 candidate feature factors(the reflectance of 10 spectral bands and 9 vegetation indices)were generated.Finally,Relief F-RF,Relief F-BPNN and Relief F-SVM were established to screen out the most effective features for soybean identification and examine the performance of the three models.Results showed that seven optimum features containing red-edge B6(740 nm),near-infrared B8(842 nm),short-wave infrared B12(2190 nm),green B3(560 nm),red-edge position(REP),red-edge normalized difference vegetation index derived from B8 and B6(NDVIre2)and enhanced vegetation index(EVI)were screened out by the Relief F-RF.The hierarchical extraction strategy was more advantageous in soybean identification and the Relief F-RF model with Kappa ranging from 0.72～0.81 performed the best and the overall accuracy was between 85.92% and 91.91%,the extraction effect was better than the previous non-layered method.The features extracted by this extraction strategy focused on field vegetation,which should have better applicability and generalization in theory.(3)Study on extraction of soybean planting area based on RGB image of UAV.Five machine-learning algorithms,i.e.,XGBoost,RF,BPNN,SVM and decision tree(DT)were adopted to screen out the optimum sliding window for soybean identification based on the UAV RGB image acquired at a height of 200 m,and the Kappa coefficient was used as the evaluation index.Then,the performance of texture features,vegetation indices and color features in soybean mapping was discussed based on the optimum window,and the machine-learning models for soybean extraction were constructed based on the optimum features.Results showed that the 40×40 pixel sliding window was more conducive for UAV soybean remote sensing identification;the Kappa coefficients of the XGBoost and RF models established based on the optimum features were both 0.88,which were 0.01,0.03,and 0.05 higher than BPNN,SVM and DT respectively,and also 0.11 higher than the results based on the original R,G,and B bands.Therefore,XGBoost and RF were more suitable for UAV soybean identification.Furthermore,the optimum feature variables had significant advantages in soybean extraction.In addition,UNet was also applied to obtain the spatial distribution of soybean based on the original three bands of R,G,and B,its Kappa coefficient reached up to 0.94,which was more prominent in UAV remote-sensing extraction of soybean.The remote-sensing recognition method for soybean in areas with fragmented farmland landscape and complex planting structures was proposed by this thesis through the Sentinel-2images and UAV images.It has achieved good results and could compensate for the lack of research on soybean extraction under such planting conditions.The results of this research can provide reference for soybean extraction under similar planting conditions,and could also provide an objective basis for local agricultural departments to carry out agricultural surveys and growth assessments.