Research On Detection And Recognition Method Of Crop Pests Based On Hyperspectral Image

Agriculture plays a vital role in the development of our country,but the frequent pests and diseases in the process of agricultural production seriously endanger the yield and quality of crops,and affect the speed of agricultural development.As an important strategy of agricultural production,pest control has a pivotal role and significance in agricultural development.The key to pest control is to obtain pest information in a timely and accurate manner.Most of the existing methods rely on manual inspection and traditional visual methods.These methods have problems such as low detection efficiency and poor recognition effect,making it more and more difficult to meet the needs of current agricultural production.Therefore,the study of a precise and efficient pest detection and identification method has important value for pest control.Hyperspectral images contain rich spectral and spatial information,and have the characteristics of unity of map and spectrum,which have become an important means of agricultural monitoring nowadays.For the problem of crop pest control,this thesis proposes a method for pest detection and identification based on hyperspectral image data and deep learning network models.Firstly,the HPDnet model is designed to detect pests by using self-coding network and anomaly detection algorithm,and it can distinguish background and pests without detailed labeling data.HPDnet extracts a variety of features of hyperspectral image data through double-branch convolutional self-coding structure,which solves the problem of missed detection in complex situations such as large area of pest occlusion and similar background,and improves the generalization ability of the network.Then,the HPInet model is designed based on parallel connected multi-resolution convolutional neural network and hyperspectral image classification algorithm,which can be used for accurate species identification of pests detected by HPDnet.HPInet fully extracts the joint spatial-spectral features of hyperspectral images using parallel multi-resolution information fusion structure to achieve increased perceptual field without reducing spatial resolution,thus enhancing spatial information feature representation.Meanwhile,by rescaling the spectral feature channels and weighted aggregation of spectral spatial information,the problem of low classification accuracy when the color shapes among pests are close is solved,and the classification capability of the network for complex data is further improved.The two models were experimentally evaluated using the dataset constructed in this thesis.In the pest detection task,the HPDnet model achieved detection AUC values of 0.975 or more.In the pest identification task,the overall recognition accuracy of the HPInet model reaches 98.62%.The experimental results show that the method in this thesis can achieve accurate detection and precise identification for pests,which can provide more help for agricultural pest control work.