Abstract:

In order to explore the effect of deep learning method on forest vegetation classification based on multi-temporal GF-1 images. This paper took Mengjiagang Forest Farm in Heilongjiang Province as the research area,took multi-temporal GF-1 images and Digital Elevation Model (DEM) as data sources,and constructed a multi-feature data set by extracting spectral features,vegetation index,texture features and topographic features,and combined with VSURF algorithm for feature optimization. At the same time,the optimized U-Net,SegNet,and DeepLab V3+models were used to classify the forest stand types,and compared with the maximum likelihood method and random forest method. The results showed as follows: 1) The classification accuracy of multi-temporal images was significantly better than that of single-temporal images; 2)Based on VSURF algorithm,16 feature variables were selected from the 97 features constructed,in which NDVI,RVI,mean,homogeneity,contrast,correlation and DEM features were retained because of their high contribution,and the other variables were eliminated,so as to avoid the "dimension disaster" to a certain extent and improve the efficiency of the model;3) Among the three depth learning methods,U-Net model had the highest classification accuracy,with an overall accuracy of 87.18%,kappa coefficient was 0.710,DeepLab V3+model followed,and SegNet model had the lowest accuracy. Constructing the optimal feature combination based on multi-temporal GF-1 images,combined with deep learning methods,has certain reference value for the classification of forest stand types.

Key words: multi-temporal GF-1 image, vegetation index, texture feature, VSURF, U-Net, SegNet, DeepLab V3+