Identification of forest insects based on machine learning

Abstract

Forest is one of the core subjects of terrestrial ecosystems, and forest insects form an important part of forest ecosystems. In recent years, the monitoring of forest insects has become particularly important as forest species diversity continues to evolve. In that case, this paper proposed a methodology to identify forest insects based on machine learning. This paper focused on three questions: for the data set, owing to the lack of existing datasets of forest insects, a delta-coder technique was used to learn the features of a few-shot dataset and generate more data (enhanced dataset) for machine learning program to learn; for the neural network, the author pre-processed the neural network by training an improved ResNet-18 network on insect section of ImageNet, then, the pre-trained ResNet-18 network was used as prototypical network and was trained by the enhanced dataset; for the effectiveness estimation of the identification model, SVM, ResNet, and VGGNet were used to do the same identifying process and compare the accuracy of this paper’s work and other three algorithms. By modifying the algorithm over and over again, a machine learning-based forest insect identification model was derived with an accuracy of 97.89%.