Research On Prediction Of Stored-Grain Pests’ Trap Counts Based On Machine Learning

In the process of grain storage,the occurrence and development of stored grain pests will lead to the loss of the quantity and quality of stored grain.Therefore,it is very important to monitor the occurrence of pests in stored grain bulks in real time and predict the development trend of pest population.At present,the main method to determine the occurrence of pests in stored grain bulks is to manually sample the bulks in stored grain bulks and estimate the development trend based on the sampling results.This is not only time-consuming and laborious,but also affects the accuracy of prediction due to the contingency of the sample.As a trap and detection device,the electronic probe trap can realize real-time monitoring of the occurrence of stored grain pests.In the ecosystem of grain bulks,stored grain pests are the main biological factors.Their life activities follow their own biological development rules,and are also closely related to temperature,relative humidity and spatial location.In this thesis,under the condition of small experimental grain bulks,the electronic probe trap is used to carry out long-term real-time monitoring of pests in grain bulks.At the same time,integrated multiple factors affecting the life activities of pests,the prediction of stored grain pests’ trap counts is carried out based on machine learning method,providing a new research method for predicting the development trend of pests population.The main research work completed is as follows:1.A stored grain pests trapping dataset is established.The occurrence and development simulation experiment of stored grain pests was carried out in wheat bulks with Sitophilus oryzae adults,one of the main storage pests,and data collection was completed with electronic probe trap.The stored grain pests trapping dataset with time windows of 5 days,10 days and 20 days were established for the subsequent feature vector and prediction model.2.The effects of temperature,relative humidity and spatial location on the stored grain pests’ trap counts were analyzed under three seasonal climatic conditions of summer,autumn and winter with three experimental configurations of 0.1,1 and 5 pests/kg.This thesis also explored the changes in the spatial distribution of pests population under the above experimental environment.Combining the above experiments and analyses,the daily average temperature,daily average relative humidity,spatial location at the trapping point and degree days representing the biological development of pests were fused based on the pests’ trap counts,and the feature vector of pests trapping data was constructed as the input of the subsequent prediction model.3.A multi-timescale pests’trap counts prediction model is proposed.The short-term prediction model is a gradient boosting decision tree model with the introduction of long-time memory features(LM-GBDT),the medium-term prediction model is a two-layer LSTM network model,and the long-term prediction model is an attention mechanism weighted LSTM model(A-LSTM),which achieved the prediction of the next 5 days,10 days and 20 days of the average trap counts of a single electronic probe trap,respectively,and the R2 scores of the three models reached 0.913,0.933 and 0.942.