UAV remote sensing estimation of wheat leaf area index using a 3D RTM model

Leaf area index (LAI) is a critical biophysical parameter for characterizing crop canopy structure and growth status, and its accurate estimation is essential for high-throughput phenotyping and precision breeding at the breeding plot scale. Unmanned aerial vehicle (UAV) multispectral remote sensing provides an efficient approach for dynamic crop growth monitoring. Such plot-scale phenotyping demands methods that remain accurate and stable across heterogeneous genotypes. Most conventional LAI inversion methods depend on empirical statistical relationships or one-dimensional radiative transfer models, which fail to adequately characterize the complex three-dimensional canopy structure of row-planted crops and ignore fine-scale structural information. This deficiency makes LAI estimation susceptible to canopy heterogeneity, growth stage variations, and varietal differences, especially for breeding populations with diverse and complex canopy architectures, resulting in difficulties in simultaneously balancing estimation accuracy and stability at the breeding plot scale. To address this issue, this study proposed a physically constrained LAI inversion framework for winter wheat by coupling a three-dimensional radiative transfer model with UAV multispectral imagery. Based on the Digital Plant Phenotyping Platform (D3P), diverse three-dimensional wheat canopy structures under variable growth conditions were realistically simulated. These simulations spanned a wide range of canopy configurations, including varying leaf angles, plant densities, and soil backgrounds, thereby capturing the structural complexity of breeding plots. A total of 10000 sets of canopy reflectance data were generated to construct a physical inversion dataset, establishing a mechanism-driven inversion framework independent of empirical statistical rules. Compared with traditional empirical methods, the proposed framework explicitly incorporated three-dimensional canopy structural features and radiative transfer physical constraints, effectively improving the robustness and transferability of LAI inversion under variable observation conditions and sensor configurations. Rather than relying purely on statistical correlations, this method directly embedded canopy architectural characteristics into the inversion process, linking spectral reflectance observations to intrinsic crop growth mechanisms. Using a multi-site, multi-growth-stage, multi-variety breeding experimental network, we collected multi-temporal UAV multispectral observations and synchronous field LAI measurements across the entire growth season at two experimental sites (Baima, Nanjing and Yangling, Shaanxi), covering multiple wheat varieties and growth stages from jointing to grain filling. A total of 2500 valid field samples were acquired for model evaluation, and the inversion performance was comprehensively compared with a classic data-driven empirical model under consistent breeding scenarios. The results showed that stable and accurate LAI estimation was achieved under different experimental conditions using the three-dimensional radiative transfer model. The overall inversion accuracy reached an R² of 0.80 with an RMSE of 0.98, thus outperforming the empirical model (R² = 0.78 and RMSE = 1.04). The consistent improvement across all test sites indicates that the physical constraints generalize beyond the specific environments. The proposed method exhibited stable estimation accuracy and excellent generalization ability across different sites, growth stages, and wheat varieties. The introduction of three-dimensional canopy structural information effectively mitigated estimation instability caused by canopy heterogeneity, while the physical constraint framework improved model interpretability. Notably, this method maintained reliable cross-site and cross-variety transfer performance without additional retraining, realizing high-throughput and accurate winter wheat LAI estimation at the breeding plot scale without massive ground training samples. This study provides valuable technical references for multi-variety parallel phenotyping monitoring and offers robust technical support for precision wheat breeding and field crop growth assessment.