A Sustainable Learning Framework: UAV-Based Oat Chlorophyll Monitoring Using Radiative Transfer Models and Deep Learning Techniques

Accurately and efficiently estimating leaf chlorophyll content (LCC) and canopy chlorophyll content (CCC) is crucial for assessing photosynthetic productivity and monitoring crop growth in agroecosystems. The unmanned aerial vehicle (UAV)-based multispectral sensors, combined with machine learning models, offer high-precision solutions for estimating crop chlorophyll. However, several challenges, such as model performance limited by ground-truth data and issues with spatiotemporal scalability and transferability, hinder the broader application of these methods. This study proposed a transferable pretraining framework (PROSAIL-StDNN), which leveraged knowledge from PROSAIL simulations through three deep learning models to improve the accuracy of monitoring oat LCC and CCC using UAV-acquired multispectral data. The results indicated that PROSAIL-StDNN achieved the highest accuracy in predicting LCC ( $R ^{2} =0.745$ , RMSE $= 4.499~\mu $ g/cm2, and RPD = 1.989) and CCC ( $R ^{2} =0.822$ , RMSE $= 30.535~\mu $ g/cm2, and RPD = 2.378). Relative to common statistical models, PROSAIL-StDNN reduced LCC RMSE by 10.11%–15.86% and CCC RMSE by 11.68%–22.30%. Additionally, the PROSAIL-StDNN model demonstrated significant spatial and temporal transferability after secondary fine-tuning with data from two additional field sites, achieving $R ^{2}$ values around or above 0.8 for both LCC and CCC. This framework offers a transferable approach for advancing UAV-based monitoring of crop physiological traits in agricultural practice.