Prediction Algorithm And Application For Area Rank Of Dendrolimus Punctatus Breakout Based On Minority Oversampling

Dendrolimus punctatus is one of the most severe pest diseases that pose threats for various forests in China.Accurate prediction methods for breakout of this pest can reduce the cost of pest control and improve the control effect as well.Traditional prediction methods based on classical machine learning algorithms can hardly present ideal effect when pest dataset are characterized with small size,imbalanced distribution and high feature dimension.Therefore,the prediction result is not satisfied with low accuracy and weak stability when predicting serious pests with less training samples.Minority oversampling can increase the quantity of available samples in model training dataset and the spatial distribution balance of sample space,which presents a new idea for forestry pest prediction.Hence,a pest prediction algorithm was proposed in this thesis upon the practical forest pest prevention and early warning requirements.The proposed method focused on solving low prediction accuracy and weak stability by using machine learning methods to establish the prediction model of pest occurrence area level with imbalanced dataset.Major research work of this paper is presented as follows.1.The dataset of Dendrolimus punctatus pest was established in this thesis.The data of pest area levels and influencing factors in four selected regions of Guangdong province were collected.The data source was acquired from the National Forestry Pest Control Station and the National Meteorological Data Sharing Center.After data pre-processeing and feature selection,the most relevant factors of pest breakout was selected,which provides research materials for the prediction of the area level of Dendrolimus punctatus.2.An improved minority oversampling algorithm(Adaptive k neighbor searching ranked synthetic minority oversampling boost,AS-SMOTEBoost)was proposed.The Synthetic minority oversampling technique(SMOTE)was improved according to sample ranking,adaptive k neighbor searching and ensemble learning.The dataset expansion and balance distribution were optimized and the accuracy and robustness of prediction model were promoted with the help of this improved algorithm.Then this algorithm was applied into prediction of the area level prediction of the pest of Dendrolimus punctatus in four regions of Guangdong province,China.3.The prediction software for Dendrolim-us punctatus was designed.The software was designed based on Tkinter module of python.A variety of machine learning algorithms were integrated into this software to realize many function such as pest prediction and feature selection,which greatly satisfies the needs of users in multiple application scenarios.In this thesis,the UCI database and Dendrolimus punctatus pest dataset were used to verify the performance of the proposed algorithm.The result showed that the improved oversampling method was superior to other oversampling algorithms with better F-measure,G-mean and ROC curve,which provides reference for pest prediction.The software realized the prediction of Dendrolimus punctatus and simplified the operation process,which presents support for related research work.