Vegetation classification based on a multi-feature Sentinel-2 time series

Abstract

To provide a statistical basis for biological characteristics, ecological characteristics, and management value of different woodlands, and to provide effective data support for rational management of forests, a forest resources survey was conducted. Vegetation classification, using a hierarchical classification, was the basis for studying the status and dynamics of forest resources with vegetation types being identified quickly and accurately by means of remote sensing. The investigated site was Wangye Forest Farm in southwest Harqin Banner, Inner Mongolia. First, according to the phenological characteristics of the vegetation, images of vigorously growing vegetation were selected to calculate an NDVI and to set appropriate thresholds to extract the vegetation. Then using the NDVI time series, spectral reflectance characteristics of 10 bands in the best time of the Sentinel-2 data as well as textural features of the first three components from a principal component analysis were selected as classification features. The vegetation types in the study area were divided into the five categories of cultivated land, grassland, evergreen coniferous forest, deciduous coniferous forest, and deciduous broadleaf forest using a support vector machine classifier. Classification results were compared with the maximum likelihood method and the method of combining NDVI time series and spectral characteristics. Results showed that the overall accuracy of vegetation classification based on Sentinel-2 time series multi-features reached 87.64%. This was an increase of 15.73% compared to the maximum likelihood method and an increase of 14.61% compared to the method of combining NDVI time series and spectral characteristics. The Kappa coefficient was 0.85, which was an increase of 0.20 compared to the maximum likelihood method and an increase of 0.18 compared to the method of combining NDVI time series and spectral characteristics. Classification accuracy for the evergreen coniferous forest (95.65%) and cultivated land (92.31%) were highly consistent with the field survey. Thus, (1) combining multiple features was helpful for improving classification accuracy; (2) temporal characteristics of NDVI greatly helped to distinguish vegetation; and (3) by using the idea of stratified classification, the vegetation could be extracted first, thereby eliminating the disturbance of non-vegetation factors and effectively improving the classification precision of vegetation types.