Extraction Of Rice Planting Area From Remote Sensing Image Based On Deep Learning

Rice is one of the main food crops in my country.The rapid monitoring and automatic extraction of its planting area are of great significance to my country’s grain production estimation and national food supply security.Remote sensing classification technology is widely used in the extraction of rice planting area,but there are problems such as insufficient extraction accuracy and automation degree,and it is urgent to establish a highly intelligent and automated method for extracting rice planting area.At present,deep learning methods have been widely used in crop image recognition and automated monitoring and other fields,which can effectively make up for the current lack of precision in extracting rice planting areas.Therefore,based on the highresolution remote sensing image data,this paper uses the improved U-Net model to conduct rice planting area extraction experiments,and compares the extraction results with traditional remote sensing classification methods;on this basis,uses the improved U-Net model.U-Net model extracted the late rice planting area in Doumen District,Zhuhai City in 2019,and analyzed the accuracy of the results.The main work and conclusions of this paper are as follows:(1)Analyzing the phenological period of rice in the study area,it is found that the growth of rice is most prosperous from mid-September to early October.The characteristics of rice in the high-resolution images of this period are the most obvious,which is the use of high-resolution remote sensing images for rice cultivation The best time for area extraction.(2)An improved U-Net model is proposed.Based on the U-Net model,this paper deepens the level to effectively learn abstract rice features from the image;at the same time,by adding the Dropout layer,the model parameters are reduced,the model complexity and over-fitting are reduced.,Improve the operating efficiency of the model.(3)Based on the improved U-Net model,the rice planting area in the experimental area was extracted,and the extraction results were compared with the extraction results of the other three traditional remote sensing classification methods.The results show that the improved U-Net model has the highest total accuracy of extracting rice planting area,reaching 97.79%;while the three traditional remote sensing classification methods of supervised classification,object-oriented classification,and decision tree classification have extraction accuracy of 91.78% and 95.14 respectively.%,95.3%,are lower than the improved U-Net model extraction method proposed in this paper.(4)Using the improved U-Net model,we extracted the late rice planting area in Doumen District,Zhuhai City in 2019,and achieved good results.The extraction results show that the late rice planting area in Doumen District,Zhuhai City in 2019 is about26,300 mu,which is similar to the 28,600 mu late rice planting area disclosed by the government that year.At the same time,the accuracy of the recognition result is 93.5%compared with the field measured data,which shows that the application of the improved U-Net model to the extraction of rice planting area on high-resolution remote sensing images is feasible,and the extraction accuracy is high.