Extraction Of Crop Planting Structure Based On GF-2 Images

The extraction of crop planting structure is the basic procedure of agricultural information monitoring.In view of the inadequacy of traditional methods of crop information acquisition,including subjectivity and large errors,it is hard to meet the requirements of agricultural production and management in the information age.In recent years,remote sensing technology are gradually being applied to agricultural monitoring areas such as crop planting structure extraction because of its various advantages including its wide range,fast data collection,to name a few.The GF-2remote sensing image has the characteristics of high spatial resolution,wide coverage,and multi wave band,making it an ideal remote sensing image for crop classification and identification.This study is carried out in Shuangyang District of C hangchun to research crop planting structure based on GF-2 image.Firstly,it selects the representative experimental area in the research area to construct the initial feature space,and explore the most suitable feature space combination for the study area;Combining the feature space and Mean Shift algorithm to segment the image,and determine the optimal segmentation scale;Based on random forest algorithm and Support Vector Machine algorithm,this study respectively establishes pixel-oriented and object-oriented crop classification models,and classifies and distinguishes the corns,rice,water,roads and other ground items(including forest,grass,and crops that take place lower percentage).According to the accuracy evaluation index,it compares and analyzes four classification models,and determines the optimal classification model in the experimental area;Finally,the optimal classification model is applied to extract the planting structure of main crops(corns,rice)in Shuangyang District in 2020.The main research contents and results are as follows:(1)By feature space optimization,this study determines the most suitable feature combination for crop classification in the study area.This study makes use of the fused GF-2 remote sensing image to extract 4 spectral features,12 vegetation index features and 8 texture features related to crop classification,and constructs the initial feature space.Evaluating the importance of 24 feature s by combining random forest algorithm,and make sure the most important 10 features(spectral features: green,red;vegetation index features: green,GI,NDVI,S RI,TCARI;texture features: Mean,Contrast,Entropy)are used to construct the most suitable feature space in the research area.(2)This study compares the segmentation effect of Mean Shift with different parameters and determines the optimal segmentation scale of the image in the study area.Mean Shift algorithm with different thresholds is used to segment,compare and analyze the effect image in the experimental area,it finds that when when the bandwidth modulus are both 12 and the threshold value is 2,it effectively avoids the situation that different land features are mixed into the same patch,and it can effectively reduce the fragmentation degree of the same feature,reduce invalid patches,so as to improve the accuracy and efficiency of the subsequent object-oriented classification model.(3)It compares and analyzes the model of crop classification,and finds out more suitable extraction methods of crop planting structure in Shuangyang District.T his study constructs the pixel-oriented and object-oriented classification models to classify crops in the experimental area respectively based on Random Forest algorithm and Support Vector Machine algorithm,and evaluates and analyzes the results of classification according to the overall accuracy and Kappa modulus.It finds that among the 4 classification models,the accuracy of the classification model based on the Random Forest algorithm is the highest.The overall accuracy is as high as0.926,while Kappa modulus is as high as 0.908.Further verification based on the plot.It finds that the accuracy of classification of the Random Forest classification model is also higher than that of the Support Vector Machine classification model,among which the object-oriented classification model based on the Random Forest algorithm has even higher accuracy,with the overall accuracy of 0.971 and Kappa modulus of0.947.Therefor,it determined that the object-oriented classification model based on Random Forest algorithm is more suitable for crop planting structure extraction in Shuangyang District.(4)The planting structure of main crops in study area based on Random Forest algorithm and object-oriented classification model is extracted and analyzed.Statistics show that the corn planting area in Shuangyang District in 2020 is about 685 km2 and mainly distributed in the mountain areas on both sides of Shuangyang river,while the rice planting area is about 163 km2 and mainly planted in the low and gentle areas near Shuangyang River Valley area and Both sides of Yinma River.The sum of the planting area of the two main crops accounts for about 51% of the area.To conclude,this study has got several theoretical and practical values in the aspects of predicting the comprehensive production ability of agricultural resources,attributing the planting structures of crops in a macro view,and maintaining the balance of crop production areas.