Research On Remote Sensing Classification Mapping Of Main Food Crops In Heilongjiang Provinc

As an important part of the field of remote sensing,agricultural remote sensing mainly includes the use of remote sensing technology to identify crop types,estimate crop area and yield,and monitor the occurrence and development of agricultural natural disasters,diseases and insect pests.Accurate and timely acquisition of crop distribution map is of great significance for regional agricultural structure adjustment,agricultural policy implementation,smart agriculture and agricultural sustainable development.In agricultural remote sensing work,classification accuracy and model construction are of great significance to improve the accuracy of crop remote sensing recognition.In this study,the main grain production areas of corn,rice and soybean in Heilongjiang Province were taken as the research area.By combining the machine learning of Google Earth Engine platform(GEE)and field investigation,the classification model was constructed by random forest algorithm,support vector machine algorithm and minimum distance algorithm,combining the original spectral characteristics,vegetation index characteristics and texture characteristics.The remote sensing classification and mapping techniques of grain crops in Heilongjiang Province were studied,and the main results were as follows:1.Compared the sample set construction strategies based on manual sampling,visual interpretation,combination of manual sampling and visual interpretation,manual sampling as training set and visual interpretation as verification set,manual sampling and visual sampling as 1:1 ratio,it is found that the classification results of the sample set constructed by using visual interpretation are better and with higher accuracy.The classification accuracy is higher than that of using only manual sample points.Therefore,using visual interpretation sample points for classification not only solves theproblem of insufficient sample quantity but also improves the accuracy.During the construction of the sample set,the training set and the verification set can be constructed by combining the manual sample points and the visual interpretation sample points.2.Based on the random forest algorithm,the importance of 40 features was scored,ranked and selected.Finally,24 features were determined,including 7 original spectral features,9 vegetation index features and 8 spatial texture features.The overall accuracy reached 95.5% and the Kappa coefficient reached 0.947.3.Based on Google Earth Engine platform,the separation models constructed by random forest algorithm,support vector machine algorithm and minimum distance algorithm are compared respectively.It is found that random forest classifier is the most stable and has the highest accuracy,and can be used to construct classification models by this method.4.By comparing the classification models constructed by random forest algorithm and support vector machine method,it is found that the random forest model is more stable and more accurate in terms of user accuracy and producer accuracy of a single crop,as well as overall accuracy and Kappa coefficient.The overall accuracy of the random forest model is about 97%,while that of the support vector machine model is 93.7%.There is a 4% accuracy gap between the two models.5.In the 11 study areas of maize,rice and soybean,the average mapping accuracy was 96.22%,94.3% and 97.93%,respectively.In conclusion,the combination of machine deep learning and field measured data can better predict the planting area and distribution of main grain crops in Heilongjiang Province.