The outbreak of conifer pests will reduce the water content and chlorophyll content of the conifer trees, cause the forest canopy color to change, and even cause the forest to die, which seriously threatens the health and safety of coniferous forest ecosystem. The remote sensing monitoring of forest canopy color change can be used to evaluate the security of forest ecosystem quickly, so the study of forest canopy color discrimination is very important. Therefore, this study selected three outbreak forest areas (Binder, Ikhtamir and Battsengel) of Erannis Jacobsoni Djak in Khentiy and Hangay province of Mongolia as the experimental areas. The investigation of canopy color change information and spectrum measurement experiment in the process of larch damage was carried out. The hyperspectral characteristics and machine learning algorithm were used to distinguish different colors of Larch canopy. Firstly, through the investigation of forest in the disaster area, the color of the canopy was divided into green, yellow, red and gray. At the same time, according to the different canopy colors of healthy and damaged trees, 66 sample trees were selected from the experimental area, and their spectral canopy were measured. Secondly, the hyperspectral characteristics such as smooth spectral reflectance (SSR), differential spectral reflectance (DSR) and smooth spectral continuous wavelet coefficient (SSR-CWC) were calculated based on the canopy spectral curve of the sample tree, and the sensitivity of these hyperspectral characteristics to different colors on the canopy is revealed by means of variance analysis. Thirdly, the sensitive features of SSR, DSR and SSR-CWC were extracted quickly by using Findpeaks function and continuous projection algorithm pattern (Findpeaks-SPA). At last, the models of different color discrimination of larch tree canopy were constructed by using the random forest classification (RF) and support vector machine classification (SVM) algorithm. And compared with the Fisher discriminant (FD) model, the accuracy of the discriminant models were evaluated. The results show that: ①In multiple wavelengths of visible light, SSR-CWC showed extremely significant sensitivity to different canopy colors. ②The sensitive hyperspectral features can be extracted effectively based on Findpeaks-SAP pattern, which reduces the number of hyperspectral features and improves multicollinearity of the model. ③SSR-CWC is the most potential hyperspectral feature to distinguish different colors on the canopy. The optimal wavelet bases of Daubechies, Biothogonal, Coiflets and Symlets are db9, bior1.5, coif1 and sym4, respectively. Among them, db9-RF (SVMC) reaches the highest overall discrimination accuracy (0.900 0). It is 0.250 0 (0.450 0) and 0.250 0 (0.100 0) higher than the SSR-RF (SVMC) and DSR-RF (SVMC) models. ④The discrimination accuracy of RF and SVMC models based on DSR and SSR-CWC is better than that of FD model, especially db9-RF (SVMC) model, which overall discrimination accuracy and kappa coefficient are 0.150 0 and 0.167 0 higher than db9-FD model, respectively. It can be seen that db9-RF (SVMC) has great potential in different color discrimination of forest canopy, which can provide important reference and practical value for remote sensing monitoring of forest pest severity in forestry and ecological security related departments.