Forest Type Classification with Multitemporal Sentinel-2 Data

Abstract

There are limited numbers of studies addressing the forest status, extents, locations, types, and composition over a large area at the regional or national levels with high spatial resolution. Sentinel-2 satellites can obtain 10-to-60m multispectral imagery with high image acquisition frequency, and thus makes it possible to improve the performance of forest mapping and forest type identification over a large area. This study presents the initial classification result of the forest/non-forest covers with 5 forest types (evergreen coniferous, deciduous coniferous, evergreen broadleaf, deciduous broadleaf, and mixed forest) by using the multitemporal Sentinel-2 data set. Based on the Google Earth Engine platform, Sentinel-2 multispectral imagery and ancillary topographic data sets over the Shangri-la, Yunnan province, China, were collected. First, the Random Forest classifier was used to derive a forest/non-forest map. Second, the forest cover was sub-classified into evergreen coniferous, deciduous coniferous, evergreen broadleaf, deciduous broadleaf, mixed forest. The overall accuracy for the forest/non-forest covers reached over 98% and declined to 75.75% for the sub classification of the forest types. Furthermore, seasonal images combine with topographic information and vegetation indices improved the classification accuracy from 68.36% to 75.75%. The study confirmed the potential of the multitemporal Sentinel-2 data to accurately delineating the forest cover and forest types at the regional scale.