Physics-Informed Transfer Learning Reduces Simulation to Reality Gaps for Winter Wheat Traits Retrieval from Hyperspectral Observations

Abstract

Accurate retrieval of crop structural and physiological traits from remote sensing data remains challenging due to limited field observations and poor cross-platform generalization of data-driven models. This study develops a physics-informed transfer learning framework to quantify the contributions of improving simulated data fidelity and increasing model complexity to retrieving winter wheat leaf area index (LAI) and canopy chlorophyll content (CCC) from hyperspectral observations. Two PROSAIL-D datasets with default and physically optimized leaf angle distributions were generated to represent different levels of simulation fidelity. Four dual-branch deep learning architectures (CNN, CNN–SE, CNN–Transformer, and CNN–SE–Transformer) integrating spectral bands and vegetation indices were pretrained on simulated datasets and transferred to real observations using progressive fine-tuning. Model performance was assessed using ground-based and unmanned aerial vehicle (UAV) hyperspectral datasets, and SHapley Additive exPlanations (SHAP) analysis was applied to interpret feature contributions. Results demonstrated that transfer learning substantially improved cross-domain generalization, while enhancing simulation fidelity provided greater performance gains than increasing network complexity. The CNN–Transformer model pretrained on physically optimized simulations achieved the highest accuracy and robustness for both LAI and CCC retrieval. At ground and UAV scales, it achieved LAI estimation accuracies of R2 = 0.55 (RMSE = 0.63) and R2 = 0.53 (RMSE = 0.62), respectively. For CCC estimation, the model obtained R2 = 0.59 at both scales, with RMSE values of 36.12 μg cm⁻2 and 37.56 μg cm⁻2 for ground and UAV observations, respectively. SHAP analysis indicated that physically optimized simulations shifted model attention toward physiologically relevant vegetation indices, whereas default simulations induced stronger dependence on unstable visible wavelengths. Physically informed simulation design combined with transfer learning effectively reduces simulation to reality discrepancies, whereas increasing deep model complexity alone provides limited improvement. The proposed framework offers an accurate, interpretable, and scalable solution for cross-platform crop trait retrieval from hyperspectral observations.

Plant phenotyping relevance

ハイパースペクトル観測から冬コムギのLAIと群落クロロフィル含量を推定する物理情報付き転移学習フレームワークを開発し、地上およびUAVデータで性能評価しており、植物形質取得・推定手法が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.