Object Detection Method Of Pine Wood Nematode Diseased Tree Based On Remote Sensing Image

Monitoring pine wood nematode diseases on remote sensing images is of great significance to the economy and the environment.However,there are problems in the monitoring process,such as erroneous detection of diseased trees and missed detection of diseased trees.(1)In autumn,the forest background environment is complex,and there are a large number of other tree species that have similar characteristics to pine wood nematode diseased trees.General target detection networks are prone to interference from noise information of other tree species when extracting features,and the interactivity between extracted feature maps at each layer is poor,making the model prone to confuse diseased trees with other tree species,leading to false detection during the detection process.(2)When acquiring images through unmanned aerial vehicles,due to changes in mountain height,diseased trees between valleys are relatively small in the image,and the model is prone to ignore the learning of these small diseased tree features,resulting in a large number of missed diseased trees,resulting in poor practical application of the detection algorithm.In response to the above-mentioned issues,this article focuses on the following two aspects of research:(1)In order to improve the diversity of training samples,this paper proposes a feature difference guided sample selection strategy.Cluster analysis and quantitative statistics are performed on diseased tree features through clustering methods.Mosaic data enhancement is performed on the images corresponding to the smaller number of features to improve the representativeness and diversity of training samples;In order to reduce noise and increase information interaction between feature maps,this paper proposes a hybrid attention feature extraction network.Firstly,the channel attention module is used to weight the feature map channels,making the model pay more attention to important channels,emphasizing important channels while reducing noise.Secondly,the feature map is upsampled and resized,and then orderly spliced by channel,Finally,input the spliced feature map into the spatial attention module to locate the region of interest,and finally orderly disassemble the resulting feature map to obtain the same feature map as the original.Experiments have proved that the feature difference guided sample selection strategy and hybrid attention feature extraction network proposed in this paper can effectively improve the accuracy and recall rate of model detection.(2)To solve the problem of missing detection of small target disease trees,this paper proposes a multi scale feature fusion search network.In this paper,we first design a feature fusion network by analyzing the dataset,and use a neural architecture search network to search this network.The feature fusion network designed is used to fuse the feature map from top to bottom,from bottom to top and jump connection,fully fusing the Semantic information and location information of the target diseased tree.At the same time,the neural architecture search network is used to automatically search the feature fusion network to obtain the feature fusion network that is most suitable for our data set,so as to improve the network’s detection ability for small target diseased trees;In order to solve the problem of one object with multiple frames,this paper proposes a hybrid weighted fusion frame mechanism,which fuses multiple prediction frames generated by the same target to obtain a new prediction frame that is closer to the target disease tree.Experiments have proved that the proposed multi-scale feature fusion network and hybrid weighted fusion frame mechanism can effectively improve the accuracy and recall rate of model detection.The designed multiple sets of comparative experiments and ablation experiments show that the method proposed in this paper has better results compared to baseline and other models,and has great potential in detecting pine wood nematode disease trees in autumn and winter,with both cost savings and good performance.