Research On Intelligent Identification Algorithm Of Spodoptera Frugiperda Larval Stages Based On Machine Learning

The spodoptera frugiperda is one of the top ten warning pests in the world,and its invasion into China has caused significant losses to domestic grain production such as corn and wheat.Accurately identifying the instar stage of the spodoptera frugiperda larvae during their developmental stage,which can provide an early warning to prevent the spread of insect damage after they mature,and provide a basis for developing scientific pest control strategies.Currently,the instar determination of spodoptera frugiperda larvae is based on manual field surveys and sampling,followed by expert identification of pest species and instar stage.This method is inefficient and subject to high subjectivity.Traditional machine learning identification methods also face challenges in achieving automated feature extraction in feature calculation.To improve the accuracy of identifying the instar stage of the larvae and reduce labor costs,this paper designs a portable spodoptera frugiperda larva image acquisition device to facilitate the collection of larval images.Additionally,to more accurately identify the instar stage of the larvae,this paper proposes a spodoptera frugiperda larva segmentation algorithm based on deep learning,which automatically calculates the geometric features of the larvae and establishes a larval instar identification model based on texture features,SIFT features,and other characteristics.The main research contents and results are as follows:(1)Collection and processing of images about spodoptera frugiperda larvae.To conveniently collect images of pest larvae,a portable image collection device was designed in this study.A total of 1376 high-definition images of spodoptera frugiperda larvae were captured using this device,and an image database of spodoptera frugiperda larvae at different instar stages was established.In addition,the YOLOv4 object detection algorithm was used to locate and extract the larvae in the images,and various image enhancement strategies were used to complete preprocessing and enrich the image dataset.(2)Research and comparison of image segmentation algorithms based on spodoptera frugiperda larvae images.By segmenting the larvae in the image,the calculation of subsequent morphological characteristics is facilitated.This study focuses on comparing traditional image segmentation algorithms with semantic segmentation algorithms in deep learning such as UNet,UNet++,and Deep Labv3+.Different semantic segmentation models were built separately using the same training set,and the segmentation effect was evaluated using metrics such as the average intersection over union(MIo U).This paper proposed optimization strategies such as scaling mapping,dual attention mechanism,and bicubic interpolation to improve the UNet++ algorithm’s edge segmentation smoothness and low accuracy in segmenting grasshopper larvae.These strategies enhance the recovery and reconstruction abilities of the target feature map,and the corresponding model is trained to effectively improve the segmentation accuracy of grasshopper larvae in images.(3)Research on the identification algorithm of the instar of spodoptera frugiperda larvae based on improved random forest.Aiming at the problems that there are too many redundant feature values in the samples,it is easy to form a complex tree structure,resulting in too large a model and a long time-consuming data processing,etc.,this paper optimized the larval geometric feature extraction process and model complexity.The difficulty of insect length extraction was reduced by methods such as bifurcation elimination and endpoint extension,and the head and tail judgment is optimized by using the gray frequency pre-judgment comparison to complete the head capsule width measurement.Additionally,the Boruta feature selection method constructs the final feature space,retaining effective features and reducing the complexity of the instar identification model.Furthermore,the optimization of the parameter search strategy using random search and grid search improved the average accuracy of instar identification.(4)Testing and result analysis of an automatic instar identification algorithm for spodoptera frugiperda larvae.Using Deep Labv3+ 、 UNet 、 UNet++ and optimized UNet++ models to segment images of spodoptera frugiperda larvae in different instar stages,which were collected by portable image acquisition devices.The test results showed that the improved MRES-UNet++ model had the highest segmentation accuracy,with PA and MIo U reaching 98.39% and 93.82%.Compared to the original UNet++model,it has increased by 2.29% and 6.16%,respectively,while also showing significant improvements in target edge smoothness.In the instar period classification process of the spodoptera frugiperda larval instar identification algorithm,the improved Boruta-DS-RF model performed the best on identifying the instars of spodoptera frugiperda larvae,achieving an average identification accuracy of 94.20%.Compared with the original random forest algorithm,the average identification accuracy is improved by 14.01%...