文献类型：期刊文献

英文题名：Physics-Driven Machine-Learning Retrieval and Uncertainty Quantification of Crop Leaf Area Index

作者：Liu, Wei[1,2] Zhu, Xiaohua[1] Yang, Suyi[3] Gao, Zhihai[4]

第一作者：Liu, Wei

通信作者：Zhu, XH[1]

机构：[1]Chinese Acad Sci, Aerosp Informat Res Inst, Beijing 100094, Peoples R China;[2]Univ Chinese Acad Sci, Sch Elect Elect & Commun Engn, Beijing 100049, Peoples R China;[3]Cent South Univ, Sch Geosci & Info Phys, Changsha 410083, Peoples R China;[4]Chinese Acad Forestry, Inst Forest Resource Informat Tech, Beijing 100091, Peoples R China

年份：2025

卷号：17

期号：23

外文期刊名：REMOTE SENSING

收录：;EI(收录号：20255119713981);Scopus(收录号：2-s2.0-105024666102);WOS:【SCI-EXPANDED(收录号:WOS:001635230700001)】；

基金：This work was funded by the National Key Research and Development Program of China (2023YFB3905804; 2022YFB3903501), Key R&D Program of Shandong Province, China (2025CXGC010113); and the Future-Star program of the Aerospace Information Research Institute (E2Z106010F).

语种：英文

外文关键词：leaf area index; neural network; genetic algorithm; uncertainty quantification; PROSAIL radiative transfer model

摘要：Highlights What are the main findings? A PROSAIL-driven, GA-optimised MLP (NN-GA) reliably retrieves crop LAI from Sentinel-2B at 10 m, achieving RMSE/R2 = 0.44/0.73 (Minqin) and 0.40/0.56 (Zhangye), outperforming the SNAP/SL2P benchmark. A staged uncertainty quantification (UQ) workflow separates physical-driver and machine-learning contributions and synthesises them to report retrieval relative uncertainties (Minqin 21.37%, Zhangye 17.31%). What is the implication of the main finding? The framework improves 10 m LAI retrieval accuracy and delivers a reproducible, end-to-end uncertainty decomposition to support confidence-aware agronomic applications. The results prioritise reductions in machine-learning stage stochasticity and recommend including uncertainty as a routine product layer to increase LAI product reliability.Highlights What are the main findings? A PROSAIL-driven, GA-optimised MLP (NN-GA) reliably retrieves crop LAI from Sentinel-2B at 10 m, achieving RMSE/R2 = 0.44/0.73 (Minqin) and 0.40/0.56 (Zhangye), outperforming the SNAP/SL2P benchmark. A staged uncertainty quantification (UQ) workflow separates physical-driver and machine-learning contributions and synthesises them to report retrieval relative uncertainties (Minqin 21.37%, Zhangye 17.31%). What is the implication of the main finding? The framework improves 10 m LAI retrieval accuracy and delivers a reproducible, end-to-end uncertainty decomposition to support confidence-aware agronomic applications. The results prioritise reductions in machine-learning stage stochasticity and recommend including uncertainty as a routine product layer to increase LAI product reliability.Abstract Leaf Area Index (LAI) is a key biophysical descriptor of crop canopies and is essential for growth monitoring and yield estimation. We present a physics-driven machine-learning framework for operational LAI retrieval and end-to-end uncertainty quantification that couples the PROSAIL radiative transfer model with a genetic-algorithm-optimised multilayer perceptron (NN-GA). PROSAIL is sampled across plausible parameter priors and spectra are convolved with Sentinel-2B spectral response functions to build a 30,000-sample training library; a GA is used to globally optimise network weights and biases. Total retrieval uncertainty is decomposed into a simulation component (PROSAIL parameter variability) and a training component (variability across repeated NN-GA trainings) and combined via the law of propagation of uncertainty. The model was developed in Minqin (modelling/testing area; entirely maize) and transferred to Zhangye (transfer/validation area; predominantly maize, with one sunflower plot). Sentinel-2B validation results were RMSE/R2 = 0.44/0.73 (Minqin) and 0.40/0.56 (Zhangye), indicating reasonable cross-site generalisation. The uncertainty split indicates physical-driven contributions of 11.42% and 11.48% and machine-learning contributions of 18.06% and 12.96%, respectively. The framework improves 10 m LAI retrieval accuracy and supplies a reproducible, per-pixel uncertainty budget to guide product use and refinement.