Mapping 30 m Fractional Forest Cover over China’s Three-North Region from Landsat-8 Data Using Ensemble Machine Learning Methods

• Published in: Remote sensing (Basel, Switzerland) Vol. 13; no. 13; p. 2592
• Main Authors: Liu, Xiaobang, Liang, Shunlin, Li, Bing, Ma, Han, He, Tao
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.07.2021
• Subjects: Algorithms; Biodiversity; Climate change; Datasets; Deserts; Environmental changes; Forests; fraction forest cover; Gaofen-2; Image resolution; Land area; Land surveys; Landsat; Landsat satellites; Landsat-8; Learning algorithms; Machine learning; machine learning algorithm ensemble; Precipitation; Remote sensing; Root-mean-square errors; Satellite imagery; Spatial distribution; Three-North region of China; Vegetation; China
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs13132592

The accurate monitoring of forest cover and its changes are essential for environmental change research, but current satellite products for forest coverage carry many uncertainties. This study used 30-m Landsat-8 data, and aggregated 1-m GaoFen-2 (GF-2) satellite images to construct the training samples and used multiple machine learning algorithms (MLAs) to estimate the fractional forest cover (FFC) in China’s Three North Region (TNR). In this study, multiple MLAs were merged to construct stacked generalization (SG) models based on the idea of SG, and the performances of the MLAs in the FFC estimation were evaluated. The results of the 10-fold cross-validation showed that all non-linear algorithms had a good performance, with an R2 value of greater than 0.8 and a root-mean square error (RMSE) of less than 0.05. In the bagging ensemble, the random forest (RF) (R2 = 0.993, RMSE = 0.020) model performed the best and in the boosting ensemble, the light gradient boosted machine (LGBM) (R2 = 0.992, RMSE = 0.022) performed the best. Although the evaluation index of the RF is slightly better than that of the LGBM, the independent validation results show that the two models have similar performances. The model evaluation results of the independent datasets showed that, in the SG model, the performance of the SG(LGBM) (R2 = 0.991, RMSE = 0.034) was better than that of the single or non-ensemble model. Comparing the FFC estimates of our model with those of existing datasets showed that our model exhibited more forest spatial distribution details and higher accuracy in complex landscapes. Overall, in this study, the method of using high-resolution remote sensing (RS) images to extract samples for FFC estimation is feasible. Our results demonstrate the potential of the ensemble MLAs to map the FFC. The research results also show that among many MALs, the RF algorithm is the most suitable algorithm for estimating FFC, which provides a reference for future research.