:Near-infrared spectroscopy was widely used as a non-destructive and fast method in agriculture.To solve the low accuracy,poor reliability and high cost of the detection of Huanglongbing,amethod of citrus near-infrared spectroscopy for Stacking Denoising Auto-encoders Combined Random Forestis proposed.The method firstly preprocesses the spectral data by multi-stage preprocessing.Then,by usingSDAEthedimensionof the pre-processed spectral datasetare reduced and deep features of the sample are extracted.Finally,the voting ensemble strategy of RF is used to realize classification and identification.In order to verify the performance of the SDAE-RF model,the near-infrared spectrum data of citrus leaves provided by a company were taken as an example.Experiments were carried out with different proportions of training sets,and compared with the identification capabilities of ELM,SWELM,SVM,BP,SDAE and RF models.Experimental results show that the SDAE-RF model performs better than other algorithms in terms of classification accuracy,algorithm stability and train time.