Pine Wilt Disease Recognition And Application Based On Deep Learning

The monitoring of pine wilt disease is the focus of China’s forestry pest control work,pine wood wilt disease spread rapidly,difficult to control,the speed of death and other problems.Since the disease was introduced to China in 1982,it has killed more than 500 million pine trees,causing extremely serious natural resource and economic losses.Therefore,the prevention and control of pine nematode trees is an urgent problem of natural resource regulation in China.At present,the methods of pine wilt tree supervision in China mainly include: ground manual monitoring,trap monitoring,remote sensing monitoring and other technologies.Manual monitoring on the ground is less efficient and very labor-intensive,and trap monitoring is mainly done by hanging traps to trap infectious sources such as pine wood nematode to achieve the monitoring effect,and the effect of this method is not stable.The use of remote sensing or drone images for monitoring is currently the most common means of supervision.Satellite remote sensing images based on spectral information have the problem of low spatial resolution,and it is difficult to identify specific tree plants on remote sensing images.Although multispectral-based drone data can obtain the spectral information of plants,there exists a higher cost of multispectral cameras,which is not suitable for wide application on a large scale.Visible light-based UAV images,although without spectral information,are cheaper to obtain than multispectral images and have higher spatial resolution.With the increasingly widespread application of deep learning methods in pest detection,the method of combining deep learning methods with visible light images has become a new direction for pest detection.In order to reach the application level,the spatial resolution of the visible images used in the experiments is low.The problems brought about by the low spatial resolution of visible images are mainly:(1)The targets of single pine trees such as horsetail pine are small and account for a small percentage of the images relative to the whole data set,which leads to far more samples sampled on the background than on the target samples when the deep learning model is sampled,which seriously affects the training of the model accuracy.(2)The dataset of diseased tree detection has complex features,and the diseased tree samples and some easily confused features are not easily separated,and these samples have large gradients,so that the model forcibly fits the abnormal samples not only cannot optimize the model parameters,but also will reduce the robustness of the model.Based on the above problems,two methods are proposed in this paper to solve the above problems,mainly including.(1)A sampling algorithm based on sampling threshold interval weighting is proposed.Firstly,the sample space is divided into simple sample intervals and difficult sample intervals,and according to the distribution law of samples,the proportion of simple samples is much larger than the proportion of difficult samples.Simple samples are the samples that the model can easily classify,and a large number of learning simple samples cannot improve the ability of the model,while difficult samples are the samples that the model is difficult to classify and are the samples that the model needs to learn.Therefore,this paper designs a weighting strategy through experiments,based on the ratio of the sample size within the current sampling threshold to the overall sample space as a weighting factor,to increase the sampling proportion of difficult samples to decrease the sampling proportion of simple samples,so as to optimize the overall sample sampling quality.(2)A loss function based on the weighting of the sample gradient distribution is proposed.The loss function gives the model optimization direction to the loss function based on the gap between the true value and the estimated value of the model.In the disease tree detection,the features are extensive,in which there are many features similar to the target samples.The difference between the estimated value of the model and the true value of these features is small,however,these samples are not the target samples and will give the wrong optimization direction to the model,and these samples we call anomalous samples.In this paper,we propose a loss function based on the weighting of the sample gradient distribution according to the gradient distribution law.The range of the gradient is between 0-1,and the number of samples gradually decreases with the increase of the gradient,which is because the number of simple samples is much larger than the difficult samples,and the closer the gradient is to 1,the more the sample size happens to increase abnormally,and this part of the sample is the abnormal sample.By dividing the gradient interval and counting the unit interval sample size,the experiment can dynamically adjust the weight of the loss function through the unit interval sample size to achieve the effect of suppressing the gradient of abnormal samples because the number of abnormal samples is more than that of difficult samples.Combining the above methods,a two-stage pine nematode tree target detection model based on sampling threshold interval weighting method and sample gradient distribution weighting mechanism is proposed in this paper.The model’s successfully improves the performance of the detection model by increasing the sampling rate of difficult samples and suppressing abnormal sample gradients.Experiments were conducted on the pine nematode tree dataset labeled in this paper,and the results showed that the proposed model obtained the best performance in pine nematode tree detection.Compared with other mainstream detection models,the detection network proposed in this paper can identify the diseased trees more accurately in the lower spatial resolution dataset.