Study On Crop Classification Method Based On GF Satellite Remote Sensing Image

Timely and accurate acquisition of crop planting structure information can provide important data support for crop growth monitoring,yield estimation and evaluation of agricultural ecological environment.The crop planting structure information obtained by means of statistical reports and sampling surveys has problems such as poor real-time performance and large errors.Since remote sensing technology has the advantages of large monitoring range and rapid data acquisition,this study applies remote sensing technology to the analysis of crop planting structure,and studies the method of crop classification based on remote sensing images.The domestic GF-1 remote sensing image has the characteristics of high spatial resolution and short revisit period,and is gradually becoming the main data source in the field of crop classification and identification.In this study,the GF-1 remote sensing image of Shenyang City,Liaoning Province was used as the data source,and the field test data was obtained through field investigation,calculation of the four vegetation index characteristics results of NDVI,EVI,NDWI and SAVI in each period of images,and the time series multi-vegetation index characteristic image was constructed.Establish a Support Vector Machine model,a Random Forest model,and a deep residual neural network(Res Net)classification model for crop planting areas,compare and analyze the classification results,and select a relatively better crop classification model.The main contents and results of this study are as follows:(1)Based on the four time-phase GF-1 remote sensing images of Shenyang City after preprocessing,four vegetation index features of NDVI,EVI,NDWI and SAVI were extracted,and the sixteen feature images obtained were used to construct a time-series multi-vegetation index iamge.Compare and analyze the time series curves of typical ground objects corresponding to these four vegetation index.According to the changes of each vegetation index in different time-phase images,the dynamic change process of ground objects can be described more accurately,which can be used for crop classification.(2)Two classification models,support vector machine and random forest,were constructed using time-series multi-vegetation index images.Combined with the experimental data collected from field investigations in the study area,a training set and a test set were constructed to extract crops in the study area,and the classification accuracy was evaluated.The results show that the overall classification accuracy of the SVM classification model is62.04%,and the Kappa coefficient is 0.4775;the overall classification accuracy of the RF classification model is 84.26%,and the Kappa coefficient is 0.7884.The classification accuracy of the RF model is high,and it is superior to the SVM model in the producer and user accuracy of various ground objects.(3)Using time series multi-vegetation index images,a classification model based on deep residual neural network was constructed to realize the classification and extraction of crops,compare and analyze the result accuracy of the classification model.The test results show that the overall classification accuracy of the Res Net classification model is 90.21%,and the Kappa coefficient is 0.8775.Compared with the SVM classification model,the overall classification accuracy is improved by 28.17%,and the Kappa coefficient is improved by 0.4.Compared with the RF model,the overall classification accuracy is improved by 5.95%.The Kappa coefficient is increased by 0.0891.The research shows that the Res Net classification model is more suitable for the extraction of rice planting areas in the study area.It is estimated that the planting area of rice in the classification results of the RF model is 1395.76km~2,which is 214.78km~2deviated from the Shenyang Statistical Yearbook data.The Res Net classification model estimates that the planting area of rice in the classification results is 1040.56km~2,which is 140.42km~2deviated from the statistical yearbook data,which is smaller than the other two classification models.This shows that the Res Net classification model is more suitable for crop classification and planting structure analysis in the study area.