A Novel Method for Mapping Moso Bamboo Forests Using Remote Sensing Data With the Consideration of Phenological Status

Precisely delineating the distribution of moso bamboo forests is critical for forestry management and regional carbon cycle research. The unique phonological characteristics (i.e., on- and off-year phenomenon) of bamboo impose difficulties in bamboo identification. This study aims to develop a new algorithm for mapping bamboo distribution using remote sensing data with the consideration of bamboo phenological characteristics. Three optical indices were proposed based on canopy reflectance retrieved from Sentinel-2 and field inventory data, including MBI (Modified Bamboo Index), BPCI (Bamboo Phenological Characteristic Index), and BPCI-2 (Bamboo Phenological Characteristic Index 2). The collaboration of these three indices with the RFE (recursive feature elimination) and XGBoost (extreme gradient boosting) methods can precisely mapping bamboo distribution and its phenological status. The model based on MBI, BPCI, and BPCI-2 outperformed the model driven by existing bamboo extracting indices, i.e., BI (Bamboo Index), YCBI (Yearly Change Bamboo Index), MCBI (Monthly Change Bamboo Index), increasing in overall accuracy (OA) by about 1.5%. Additionally, proposed indices were calculated using the data synthesized from Sentinel-1 SAR (synthetic aperture radar) imageries by the CycleGAN (cycle-consistent adversarial network) method under the condition without cloudy-free Sentinel-2 data available to fill the time series data gaps. The performance of model based on augmented data improved notably in comparison with the model driven only by indices from original optical images, with the identification accuracy for on- and off-year bamboo samples over 96%. The generated moso bamboo distribution map aligns well with forestry inventory data in terms of both area and spatial distribution. The proposed indices are less sensitive to terrain than existing bamboo extracting indices. This merit is valuable for better mapping bamboo forests, which are mostly distributed in mountainous areas.