Research On Detection Method Of Grassland Degradation Indicator Grass Species(Stellera Chamaejasme Linn) Based On Deep Learning

Due to climate change and human factors,grassland degradation occurred in Three Rivers Source area.At present,most grassland workers use manual methods such as visual inspection and other artificial methods or remote sensing technology to carry out the macroscopic evaluation of grassland.After a long period of research,grassland experts have discovered the emergence of grass species indicative of grassland degradation is an important sign of grassland degradation.Therefore,it is simpler and more convenient to use the detection technology of degraded indicator grass species for early warning of grassland degradation.This paper takes the degraded indicator grass species Stellera chamaejasme Linn as an example,and introduces the research background content and significance of this subject.This paper combines the research status of deep learning and object detection at home and abroad,basic concepts of Convolutional neural networks,traditional object detection algorithm and object detection algorithm based on Convolutional neural networks.It mainly introduces the YOLO series algorithm,the SSD algorithm and the RCNN series algorithm.The representative grass species for grassland degradation have been studied,among them are Achnatherum inebrians(Hance)Keng,Achnatherum splendens(Trin)Nevski,and Stellera chamaejasme Linn.Because Stellera chamaejasme Linn is bright and distinct from the grass background,Stellera chamaejasme Linn is used as the research of this subject.Aim at the object detection work for grassland degradation indicator grass species(Stellera chamaejasme Linn),the first step is to establish a object detection task data set for grassland degradation indicator grass species(Stellera chamaejasme Linn).The feasibility and pros and cons of the Convolutional neural network on this data set,and then discuss the performance differences and reasons of the YOLOv3-SPP algorithm model before and after the self-built data set cutting technology is applied.Finally,choose the YOLOv3-SPP,SSD,Faster RCNN algorithm to carry out the contrast experiment of the Convolutional neural network on the degraded indicator grass species(Stellera chamaejasme Linn)data set.The experimental result is: the accuracy rate of the YOLOv3-SPP model: 90.2%,Recall rate: 93.4%,AP value: 92.5%,SSD model accuracy rate: 74.3%,recall rate: 75.2%,AP value: 74.43%,Faster RCNN model accuracy rate: 89.0 %,recall rate: 91.8%,AP value: 91.3%.According to the results of comparative experiments,based on the algorithm model of Faster RCNN,a more suitable detection model of Faster RCNN-F is proposed.The main improvement points are as follows: Firstly,the feature extraction network based on encoder and decoder type is used.Secondly,amend the SENet channel attention mechanism to spatial attention mechanism.Thirdly,using the KL Loss Loss function and Softer NMS to redundant computation.The accuracy rate of Faster RCNN-F detection model was 93.0%,the recall rate was 96.8%,and the AP value was 94.8%.Finally,the detection model before and after the improvement of the comparison work,analysis of their differences.Finally,the Faster RCNN-F detection model is used as the core to establish a object detection system for degraded indicator grass species(Stellera chamaejasme Linn).