基于圖像特征與機器學習的烤煙煙葉產(chǎn)量預測方法

打開文本圖片集

中圖分類號：S572；S126 文獻標識碼：A文章編號：1007-5119（2025）04-0087-11

Abstract： To explore the feasibilityof predicting flue-cured tobacco yield based on RGB images in combination with machine learninganddeepleamingalgorithms，afieldexperimentwascariedoutusingflue-curedtobaco Zhongchuan208.RGBimagesof flue-cured tobacco wereobtaied 25days aftertopping using drones.Color，textureandshapeoftheimages were extracted，totaling 35 features.Featuresselection were performed using therandom forest algorithm，and a yield prediction model was constructed using seven machineleaing algorits （BPNN，GA-BPNN，ELM，PSO-ELM，SVR，GA-SVR，RF）andonedeeleaingalgorithm（1DCNN）. The results showed as the follows：The accuracy （R2=0.970） and generalization ability （R2=0.817） of the random forest predictionmodelestablishedbythecombinationoffeatures（color，shape，andtexture）obtainedthroughtherandomforestalgorithm are higher hanthoseof te othersix machinelearning modelsand theconvolutional neuralnetwork model.The predictedyieldare in goodaccordance withthe measuredvalues.Insummaryweconstructedatobaco yieldpredictionmodelthroughthecombinationof feature selection and the random forest algorithm and provided novel tools for tobacco yield prediction.

Keywords： flue-cured tobacco; image features; machine learning;1Dconvolutional neural network; yield prediction

作物產(chǎn)量是農(nóng)業(yè)生產(chǎn)過程中最終的數(shù)量特征[1]傳統(tǒng)產(chǎn)量測定往往需要采收后才能統(tǒng)計，無法及時反饋生產(chǎn)，指導規(guī)劃，而產(chǎn)量預測可為作物管理和最終產(chǎn)量估計提供參考。（剩余17165字）