Objective To explore the classification effect of deep learning models on forest vegetation using multi-temporal Sentinel-2A/B images.

Method In this study, based on the multi-temporal Sentinel-2A/B images and Digital Elevation Model (DEM) in Mengjiagang Forest Farm in Heilongjiang Province, the JM distance of each forest category was used to determine the best single-phase. The characteristics of multi-temporal vegetation index and red edge index (DVI, mNDVI, CIred-edge, NDre1) were analyzed. Support vector machine and optimized U-Net model were used to carry out classification experiments on single-phase + DEM and single-phase + DEM + multi-temporal vegetation index respectively.

Result (1) On the basis of single-phase + DEM, when adding multi-phase vegetation index, the accuracy of U-Net model was 76.37%, which was 5.7% higher than that of single-phase + DEM; (2) The accuracy of U-Net model was higher than that of support vector machine. In addition, the deep learning U-Net model could avoid the "salt and pepper" phenomenon, and the classification results were more delicate.

Conclusion Based on multi-temporal Sentinel-2A/B images, the vegetation index and red edge index time series characteristics are constructed, and the U-Net model can improve the classification accuracy of forest types to a certain extent.