Research On Detection Method Of Grassland Rat Hole Based On Low Altitude Remote Sensing Image

Grassland is the foundation for the development of animal husbandry in China.Due to various factors such as weather changes,increased livestock carrying capacity,and local degradation of grasslands,rodent disasters occur frequently.Grassland rodent infestation poses a serious threat to biodiversity conservation,the production and living of farmers and herdsmen,and the sustainable development of animal husbandry.The number of rat holes,as one of the indicators for monitoring the degree of harm caused by rodents,has the advantages of simplicity,speed,and intuition.The combination of drone remote sensing and target detection technology has become a research hotspot in the field of rodent pest monitoring.The thesis uses the Da Jiang Yu Mavic2 unmanned aerial vehicle and takes the grassland in Xilamuren Town,Wuchuan County,Hohhot City,Inner Mongolia as the research area to obtain low altitude remote sensing images of grassland mouse holes.Machine learning and deep learning datasets are constructed,and traditional feature extraction algorithms and deep learning algorithms are used to carry out mouse hole target detection research.The thesis starts from improving the robustness of the model and conducts research on small mouse hole targets.The main content is as follows:(1)Construct machine learning and deep learning mouse hole datasets.Using unmanned aerial vehicles to collect summer and winter mouse hole images on site,and combining the actual situation of the mouse hole captured in real scenes,including blurring,dim light,and occlusion,the dataset was expanded using methods such as scale normalization,rotation flipping,gamma correction,motion blur,and mapping.A machine learning positive and negative sample dataset and a labeled deep learning dataset were constructed.(2)Construct a mouse hole detection model based on machine learning.Extract mouse hole features using three traditional manual feature extraction algorithms,HOG,LBP,and HAAR,and train them into a classifier.Use three types of trained weights to detect rat holes,mouse holes,blurred and occluded mouse holes,and analyze and compare experimental results.The experimental results indicate that traditional manual feature extraction algorithms are unable to obtain deeper semantic features,especially in mouse holes with small proportions or blurred occlusion in the entire image,making it difficult to accurately locate them.The necessity of constructing a deep learning mouse hole target detection model is given.(3)A DSE-YOLOv5s mouse hole detection model was proposed.Analyze the problems of the original YOLOv5s model in identifying rat holes and improve it.Improve the model from three aspects: Backbone backbone network,Neck layer,and Head detection head.Introduce deformable convolution to recover shallow lost information;SIoU Loss function is used instead of CIoU Loss function,and angle penalty term is introduced in prediction box detection to improve detection accuracy and speed up Rate of convergence;Improve the attention mechanism to focus on important features,address the issue of insufficient attention to small target information,and optimize feature extraction capabilities.In the experimental section,five attention mechanism models were first constructed to verify the effectiveness of the improved attention mechanism model.Secondly,ablation experiments were conducted on three improved methods to verify their effectiveness.The mAP,P,and R of the improved algorithm are 98.3%,97.9%,and 97.5%,respectively,which are 3.8%,3.6%,and 4.1% higher than the original YOLOv5s algorithm.Finally,compared with mainstream object detection algorithms such as SSD,Faster Rcnn,and YOLOv4,the mAP values increased by 30.71%,13.67%,and 21.63%,respectively.The experimental results show that the DSE-YOLOv5s model has significant improvements in accuracy and real-time performance.(4)Designed and implemented a visualization platform for mouse hole detection.Build a GUI interface using PyQt5,design graphical interface controls and connect slot functions to achieve weight selection,weight initialization,image detection,video detection,and other functions,and complete the visualization of mouse hole detection results.