@article{QI2026115115,
title = {Retrieving fine-scale leaf and soil spectral properties from canopy reflectance with differentiable 3D radiative transfer},
journal = {Remote Sensing of Environment},
volume = {333},
pages = {115115},
year = {2026},
issn = {0034-4257},
doi = {https://doi.org/10.1016/j.rse.2025.115115},
url = {https://www.sciencedirect.com/science/article/pii/S003442572500519X},
author = {Jianbo Qi and Jifan Wei and Mengqi Xia and Donghui Xie},
keywords = {Radiative transfer, Differentiable modeling, 3D, LESS},
abstract = {Accurate estimation of leaf spectral properties or biochemical parameters from canopy-level reflectance has long posed a significant challenge in quantitative remote sensing due to the inherent variability and complexity of canopy structures. This study presents 3D-Diff, an innovative differentiable three-dimensional (3D) radiative transfer approach for retrieving leaf and soil spectral properties using remote sensing imagery with known 3D canopy structures. Implemented within the LESS (LargE-Scale remote sensing data and image Simulation) model, 3D-Diff employs computer graphics-derived differentiable modeling to compute reflectance derivatives relative to scene parameters during forward simulation. A gradient-based optimization algorithm then minimizes discrepancies between simulated and observed images to estimate scene parameters. Validations using both virtual experiments and real measurements demonstrated robust performance across varying parameter quantities. Notably, the results show that the reflectance/transmittance of all scene parameters were accurately estimated with maximum RMSE of 0.015 for virtual experiments, including non-directly-observed elements, and maximum RMSE of 0.096 for real measurements. Spatial resolution significantly affected accuracy, with relative RMSE values of 0.07 % (0.05 m) and 1.4 % (2.5 m) for a dense virtual forest scene. Although computationally slower than analytic models, this work establishes differentiable radiative transfer as a promising framework for fine-scale vegetation mapping, enabling simultaneous multi-parameter retrieval while maintaining physical interpretability.}
}