A spatial downscaling framework for vegetation optical depth based on geographically weighted stacking ensemble learning

Vegetation Optical Depth (VOD) is an important microwave remote sensing parameter that reflects vegetation canopy water content and aboveground biomass. However, the relatively coarse spatial resolution of existing VOD products restricts their application in regional ecological and environmental studies. To generate high-spatial-resolution VOD data, this study selected the Yangtze River Economic Belt as the study area and used multi-source high-resolution variables related to vegetation, meteorology, and topography as model inputs. Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) were employed as base learners, while Geographically Weighted Regression (GWR) was adopted as the meta learner to develop a Geographically Weighted Stacking Ensemble (GWSE) downscaling model. Using this framework, VOD data were downscaled from 0.1° to 0.01° spatial resolution. The results indicate that the GWSE model achieved better downscaling performance than the individual RF, XGBoost, and LightGBM models, with an R 2 of 0.702, an RMSE of 0.180, and an MAE of 0.124. After downscaling, the spatial correlations between VOD and NDVI, EVI, LAI, and GPP were enhanced. At the significance level of P < 0.001, the correlation coefficient increased from 0.362 to 0.442 for NDVI, from 0.210 to 0.240 for EVI, from 0.305 to 0.424 for LAI, and from 0.277 to 0.362 for GPP. In addition, the monthly mean VOD time series before and after downscaling remained highly consistent across different land cover types and showed significant positive correlations ( P < 0.001). The strongest correlations were observed in grassland, cropland, and the overall vegetation cover area, with correlation coefficients of 0.993, 0.992, and 0.997, respectively. Overall, this study provides a feasible approach for generating regional high-resolution VOD products and offers data support for vegetation dynamics monitoring and ecological environment assessment. The generated high-resolution VOD data can further improve the understanding of vegetation responses to environmental changes and support ecosystem monitoring at regional scales.