The Research On Weed Identification In Rice Field Based On Deep Learning

Rice is a field crop and one of the important grain economic crops in China.It is widely cultivated and produced.In the process of rice cultivation,weeds are one of the important factors affecting rice production.Farmland weeds mainly refer to those non-artificially grown herbaceous plants that grow in large areas in farmland areas.Therefore,effective prevention and control of weeds in rice farmland is of great significance to ensure the yield and quality of rice.At present,the most widely used method of weed control in China is chemical control.As the name implies,chemical control uses chemical pesticides to spray weeds.While eliminating weeds,it will also affect the growth of rice.Long-term application of chemical pesticides will cause farmland pollution and harden the soil.Reasonable and effective use of pesticides and precise spraying of weeds have become an effective method to solve this problem.By spraying the areas with a large number of weeds in a focused manner,the areas with a small amount of weeds or no weeds are sprayed with a little or no spray.To achieve the goal of precise application of weeds,improve the efficiency of pesticide use.The focus of this study is to identify and classify the rice weeds in the farmland,and then generate a simple grid map of weed distribution.Based on extensive reading of domestic and foreign papers,this paper uses three semantic segmentation models to identify and classify remote sensing images of rice weeds captured by drones,and compares and analyzes the grid maps generated by these three algorithm models.The main work of this article is as follows:(1)The drone was used to shoot at a height of 30 m,and the image was stitched based on position information and color texture features after the acquisition,to form a low-altitude orthophoto image of the field block.Cut into many small pictures,and label the small pictures to divide the training set and test set.(2)The three semantic segmentation models used in this paper are FCN,U-Net and Seg Net.By using these three models to identify rice weed images and obtain their respective unified evaluation indicators,the pixel accuracy(Pixel Accuracy)of the three models is 88.8%,89.4%,84.5%;Mean Pixel Accuracy)were 69.3%,82.2%,and 77.0%;the average crossover ratio MIo U(Mean Intersection over Union)was 61.4%,68.8%,64.0%,and frequency weighted crossover ratio(Frequency Weighted Intersection over Union)respectively 82.4%,81.1%,76.8%..After analysis and comparison,it can be concluded that the recognition effect based on the Seg Net model is the best.Its pixel accuracy PA is 93.5%,the average pixel accuracy MA is 82.2%,the average crossover ratio MIo U is 62.5% and the frequency weighted crossover ratio FWIo U is 81.1%.(3)Splice the recognition maps obtained from the three models to generate the recognition map of the whole farmland,and then divide the three recognition maps into a checkerboard to generate a simple weed distribution grid map corresponding to the three models.Combined with the above,the evaluation coefficient indexes of the three model effects are summarized and analyzed.The research results of this paper show that the research method of remote sensing images based on deep learning model can effectively reflect the difference between rice and weeds,and obtain higher classification accuracy.Among the models that are also deep learning semantic segmentation,the Seg Net model does not matter Whether the accuracy or the recognition effect is the best compared to the other two,provides a basis for the subsequent generation of a prescription map for precise application of medicine,and provides a basis for decision-making for the precise application of plant protection drones.