Study On Segmentation Method Of Cotton Leaf Mite Hazards Based On Thermal Infrared Image

Cotton is the main cash crop in China,and cotton spider mite is a major pest in cotton production in China,which will cause cotton leaves to wither and fall off,seriously affecting cotton yield.At present,the investigation of cotton leaf mites is mainly carried out manually,which is not only time-consuming and laborintensive,but also faces problems such as point to area,strong subjectivity,and poor timeliness,making it difficult to achieve accurate monitoring of large-scale cotton leaf mites.This paper uses traditional image segmentation methods based on Open CV and four semantic segmentation network models based on deep learning,U-Net,FCN,Deep Lab v3+,and PSPNet,to segment thermal infrared images captured by drones,achieving the goal of identifying the harm of cotton leaf mites.The main research content of this article is as follows:(1)Traditional image segmentation methods based on Opencv.Segmentation of thermal infrared images captured by drones is performed,and corresponding evaluation indices are obtained by calculating accuracy,misclassification rate,and missed segmentation rate.In the segmentation of leaf mite infested areas in cotton fields,when the number of cotton overlapping areas is single plant,the accuracy rate is 75.77%,the missegmentation rate is 1.66%,the missed segmentation rate is 9.77%,and the time consumption is 1.08 seconds.When the number of cotton overlapping areas is multiple plants,the accuracy rate is 67.33%,the missegmentation rate is 5.45%,the missed segmentation rate is 12.88%,and the time consumption is 1.33 seconds.the segmentation accuracy of a single plant is higher,the error segmentation rate and missed segmentation rate are lower,and the time used is shorter,resulting in better results.(2)A deep learning based image semantic segmentation model.Four semantic segmentation network models,U-Net,FCN,Deep Labv3+,and PSPNet,were constructed and trained and tested using thermal infrared datasets.The accuracy rates of the four models were 84.4%,80.1%,81.2%,and 86.1%,respectively;The average pixel accuracy is 63.8%,59.0%,61.6%,and 67.3%,respectively;The average crossover ratio was 54.4%,56.0%,56.5%,and 61.2%,respectively;The recall rates were 60.9%,59.6%,60.2%,and 62.3%,respectively.After comparison,the PSPnet model has the best performance.The research results of this article show that using a semantic segmentation network model based on deep learning for the segmentation of unmanned aerial vehicle thermal infrared images can distinguish between normal cotton and cotton areas affected by leaf mites,thereby improving the accuracy of segmentation.The PSPNet network model outperforms other models in terms of identification performance and accuracy,and its identification results can provide technical support for precise prevention and control.