【目的】目前关于林型识别的研究区域主要为小范围森林区域和林场,为了探究较大范围内林型的识别方法,本研究使用Sentinel-2光学遥感影像、森林资源二类调查数据、数字高程模型(DEM)和Sentinel-1雷达遥感影像数据建立林型识别模型。【方法】以淳安县作为研究区域,针对淳安县各个小班中的7种林型:毛竹(Phyllosstachys edulis)林、茶树(Camellia sinensis)林、山核桃(Carya cathayensis)林、杉木(Cunninghamia lanceolata)林、马尾松(Pinus massoniana)林、阔叶混交林、其他硬阔林进行识别。研究采用分层的方法对林型进行识别,整体分为3层。第1层使用RF算法建立林地与非林地识别模型;第2层对林地数据进行树种结构识别,分别使用随机森林(random forest, RF)、极端梯度提升(eXtreme gradient boosting, XGBoost)和 轻量级梯度提升机(light gradient boosting machine, LightGBM)方法建立不同模型并对比分析实验结果;第3层将树种结构细分为林型。【结果】第1层RF林地与非林地识别模型总体精度为98.08%;第2层树种结构识别模型中对比了3个模型不同特征组合下的性能,其中LightGBM模型总体精度最高,达到81.43%;第3层模型对林型进行识别,基于所有特征结合雷达遥感因子建模的情况下,LightGBM模型精度为84.51%,经递归特征消除法(recursive feature elimination, RFE)选择特征后,最优精度为83.21%。【结论】通过各个模型的特征重要性图发现,光学遥感中的绿光、红光、近红外波段和红边波段,以及DEM提取的地形因子对研究区域小班林型识别有较好的效果,而Sentinel-1雷达中提取的自变量对林型的识别没有特别明显的帮助。

【Objective】 Currently, forest type identification research is mainly focused on small forest areas and forest farms. To explore the forest type identification method across a large range, this study used Sentinel-2 optical remote sensing imagery, forest resource survey data, a digital elevation model (DEM), and Sentinel-1 radar remote sensing image data to establish tree species identification model.【Method】 The research area was defined as Chun’an County, where seven forest types were modeled separately, including Phyllosstachys edulis forest, Camellia sinensis forest, Carya cathayensis forest, Cunninghamia lanceolata forest, Pinus massoniana forest, broad-leaved mixed forest, and other hard broad-leaved species. Models were divided into three layers. In the first layer, the RF algorithm was used to establish the identification model of forested land and non-forested land. In the second layer, forest structure was identified from forested land. The RF, XGBoost and LightGBM methods were used to build various models and analyze the experimental results. In the third layer, the forest structures were further divided into forest type.【Result】 The overall accuracy of the first-layer model based on the RF algorithm dividing samples into forested and non-forested land samples was 98.08%. In the second layer (i.e., forest structure recognition model), the performance of the three models under various feature combinations were compared. It was found that the LightGBM model had the highest overall accuracy of 81.43%. In the third layer, the performance indicators for seven forest type models were compared; based on the combination of all features and radar factors, the overall accuracy of the LightGBM model was 84.51%, after feature selection by the recursive feature elimination algorithm, the optimal accuracy was 83.21%.【Conclusion】 The green, red, near-infrared and red-edge bands from optical remote sensing imagery, and terrain factors from a DEM are effective in identifying the forest type. However, independent variables extracted from Sentinel-1 radar do not provide significant help to identify forest type.