Recognition Of Aconitum Leucostomum Based On UVA Image And Deep Learning Method

As an important pastoral area in China,Xinjiang has diverse and widely distributed grassland types.However,due to the influence of climatic factors and unreasonable utilization,the types and areas of grassland poisonous grass are increasing year by year,which seriously restricts the healthy and sustainable development of animal husbandry.Among them,the poisonous grass Aconitum leucostomum is particularly harmful in the Yili area.This study is the first time to use UAV remote sensing data and deep learning technology to identify the poisonous grass Aconitum leucostomum at the individual plant level.The study on the identification of Aconitum leucostomum based on convolutional neural network and the optimization of model parameters to improve the identification accuracy is of great significance for the accurate and rapid identification of poisonous grass,detection of grassland degradation and protection of grassland ecosystem.(1)During the training and testing process of Aconitum leucostomum recognition based on the VGG-16 backbone network based on the Faster-RCNN algorithm,the recognition accuracy of the model can be improved by changing the training parameters such as the anchor frame size and learning rate.When the anchor box scale is set to a candidate box of size(8,16,32)and the learning rate is 0.002,the obtained m AP is 45.67% of the results identified by the original parameter training,which increases to 67.24%.The effectiveness of this method to identify Aconitum leucostomum in natural environment is demonstrated.(2)Based on the Faster-RCNN algorithm and the SSD algorithm,two residual networks Res Net50 and Res Net101 are used as the backbone network to compare and analyze the advantages and disadvantages of the two algorithms and the effect of identifying Aconitum leucostomum.The results show that the recognition rate of the Faster-RCNN algorithm is better than that of the SSD algorithm,but the SSD algorithm is superior in the recognition speed.Faster-RCNN outperforms SSD in accuracy for both backbone networks with different depths.Faster-RCNN＿Res Net50 has the highest accuracy with a m AP value of 64.74%,and SSD＿Res Net50 has the lowest accuracy with a m AP value of 48.70%.(3)The Mask-RCNN algorithm based on instance segmentation takes two residual networks Res Net50 and Res Net101 as the backbone network for identification.By comparing and analyzing the loss value,evaluation performance index and identification effect of the two backbone networks,the results show that the identification accuracy map of the backbone network based on resnet50 is higher than resnet101,and the m AP is 66.0% and 65.3% respectively.Experiments show that the object detection(Faster-RCNN)and instance segmentation algorithm(Mask-RCNN)based on convolutional neural network used in this paper have good performance in the recognition accuracy and recognition effect of Aconitum leucostomum at the individual plant level in UAV images,especially the algorithm with residual network Res Net50 as the backbone network and the algorithm of changing the anchor frame size can realize the rapid and accurate automatic recognition of Aconitum leucostomum.in practical application,It can provide technical support for remote sensing monitoring of Aconitum leucostomum.