Forest canopy closure estimation in mountainous southwest China using multi-source remote sensing data.

Forest canopy closure (FCC) is an important biological parameter to evaluate forest resources and biodiversity, and the use of multi-source remote sensing synergy to achieve high-accuracy estimate regional FCC at low cost is a current research hotspot. In this study, Shangri-La City, a mountainous area in southwest China, was considered as the research area. The satellite-borne LiDAR ICESat-2/ATLAS data were used as the main information source. Combined with 54 measured plot data, the improved machine learning model of the Bayesian optimization (BO) algorithm was used to obtain the FCC in the footprint-scale ATLAS footprint. Then, the multi-source remote sensing image Sentinel-1/2 and terrain factors were combined to perform regional-scale FCC remote sensing estimation based on the geographically weighted regression (GWR) model. The research results showed that (1) among the 50 extracted ATLAS LiDAR feature indices, the best footprint-scale modeling factors are Landsat_perc, h_dif_canopy, asr, h_min_canopy, toc_roughness, and n_touc_photons after random forest (RF) feature variable optimization; (2) among the BO-RFR, BO-KNN, and BO-GBRT models developed at the footprint scale, the FCC results estimated by the BO-GBRT model were the best ( = 0.65, RMSE = 0.10, RS = 0.079, and = 79.2%), which was used as the FCC estimation model for 74,808 footprints in the study area; (3) taking the FCC value of ATLAS footprint scale in forest land as the training sample data of the regional-scale GWR model, the model accuracy was = 0.70, RMSE = 0.06, and = 88.27%; and (4) the ² between the FCC estimates from regional-scale remote sensing and the measured values is 0.70, with a correlation coefficient of 0.784, indicating strong agreement. Additionally, the average FCC is 0.50, predominantly distributed between 0.3 and 0.6, comprising 68.43%. These findings highlight the advantages of mountain FCC estimation using ICESat-2/ATLAS high-density, high-precision footprints and the fact that small-sample estimation results at the footprint scale can serve as training data for the regional-scale GWR model, offering a reference for low-cost, high-precision FCC estimation from footprint scale to regional scale.