| Rice is very important food crop in the world and one of the staple foods of most residents in our country.The use of drones to obtain images of rice planting areas,based on image analysis,allows the government to know the progress of rice planting,area and other agricultural information in a timely manner,which is of great significance for ensuring food security.The traditional methods of acquiring rice planting information in my country are mainly manual measurement and satellite remote sensing monitoring.The statistical method of manual survey is low efficiency and long cycle,and the accuracy is easily interfered by various factors,it can’t meet the demand of real-time control of accurate agricultural information.Satellite remote sensing is often limited by the impact of weather,return period,image resolution,etc.,and cannot meet the needs of real-time monitoring of crops.Due to the large differences in texture and spectral characteristics of rice planting areas at different times and locations,traditional machine learning methods have poor recognition results and limited expression capabilities,and manual visual interpretation requires a lot of work and is not efficient.Therefore,this paper proposes to use drones to obtain high-definition images of rice planting areas,and use Mask R-CNN instance segmentation model and U-Net semantic segmentation model to quickly extract rice planting areas,at the same time,compared with the traditional machine learning support vector machine and ISO clustering,the advantages and disadvantages of various extraction methods are obtained,find the best extraction method for rice planting areas,and improve the accuracy and efficiency of rice planting area identification.The main research contents and results are as follows:(1)The method of obtaining images of rice planting areas by unmanned aerial vehicles is studied.The images of rice planting period,rice growth period and harvest are collected by unmanned aerial vehicles,and these images are analyzed,the features of UAV Image ground objects in shape,spectrum and texture are obtained.Through the study of rice planting characteristics and image characteristics of rice planting area,in order to make more accurate sample data set and improve the accuracy of the model.(2)Combining with the image characteristics of rice planting areas,on the basis of studying traditional machine learning and deep learning methods,this paper mainly studies and designs the extraction method of rice planting area based on Mask R-CNN instance segmentation and the extraction method of rice planting area based on U-Net semantic segmentation,specifically in sample making,model optimization training,model Application and accuracy evaluation methods are studied.(3)Relevant experiments and result analysis are carried out.Taking UAV images in the study area in June and August as research data,the extraction experiment of rice planting area was carried out by using Mask R-CNN model,UNet model and traditional machine learning method.In the experimental results,the accuracy of Mask R-CNN instance segmentation model is relatively high,the precision is 0.8482,the recall is 0.9558,and the area relative error of extraction result is the smallest,only 4.75%.Compared with U-Net semantic segmentation model and traditional machine learning model,Mask R-CNN model has the best effect in extracting rice planting area.Therefore,the method of extracting rice planting area based on Mask R-CNN model and UAV image is recommended for the general survey of rice. |
