Correcting confounding canopy structure, biochemistry and soil background effects improves leaf area index estimates across diverse ecosystems from Sentinel-2 imagery

Abstract

High-spatiotemporal-resolution leaf area index (LAI) data are essential for sustainable agro-ecosystem management and precise disturbance detection. Previous LAI products were primarily derived from satellite data with limited spatiotemporal or spectral resolutions, which could be overcome with the use of Sentinel-2. While hybrid methods that integrate PROSAIL simulations with machine learning offer advantages in extracting high-spatiotemporal-resolution LAI from Sentinel-2, they still face challenges due to confounding factors related to canopy structure, leaf biochemistry, and soil background. To reduce impacts of these confounders, we developed an efficient hybrid method for Sentinel-2-based LAI retrieval. Our approach consists of random forest models trained on simulated datasets generated by PROSAIL-5B with two refinements: variable canopy fraction of fully senescent leaves (FS) and soil bidirectional reflectance factor (BRF) simulated by Brightness-Shape-Moisture (BSM) model. We corrected canopy BRF using near-infrared reflectance of vegetation (NIRV) and vegetation cover within mixed pixels (VC). For validation, we used ground measurements across different vegetation types from the Copernicus Ground Based Observations for Validation (GBOV) and Korea flux (KoFlux) sites during 2019–2023. Our results showed that coupling BSM and FS with PROSAIL-5B simulations improved hybrid LAI estimates, reducing RMSE by 10.8%–73.8%. Utilizing NIRV and VC to correct canopy BRF better quantified LAI in most vegetation types, with RMSE reduced by 15.3%–64.8%. Our hybrid method showed robust agreement with ground validation data from GBOV (R2 = 0.88, RMSE = 0.71) and KoFlux (R2 = 0.80, RMSE = 0.75). Overall, our method (R2 = 0.58–0.93, RMSE = 0.04–0.83) outperformed both the benchmark Sentinel Application Platform (R2 = 0.11–0.85, RMSE = 0.28–1.67) and data-driven (R2 = 0.09–0.85, RMSE = 0.29–0.93) algorithms in producing precise seasonal LAI data at finer resolutions. Our findings underscore the potential of the proposed hybrid method for high-spatiotemporal-resolution LAI retrieval across diverse ecosystems.