Improved chlorophyll and water content estimations at leaf level with a hybrid radiative transfer and machine learning model

Accurate and robust quantifications of leaf chlorophyll and water contents facilitate a better understanding of plant water and nutrient needs. Besides simplified spectral indices, other existing methods primarily use machine learning models to estimate plant traits from spectroscopy data. A powerful machine learning model usually relies on a large number of training samples. However, in-situ training data sampling in agriculture is often constrained by time and workforce. To cope with the issue of limited in-situ samples, this study proposed and evaluated a new hybrid method (Spiking-Hybrid) that coupled mechanistic model simulation with machine learning to estimate leaf chlorophyll a + b content (Cab) and equivalent water thickness (Cw). In addition, Spiking-Hybrid leveraged a ‘spiking’ technique that added in-situ data with extra replications into the simulated data to enhance modeling performance. We evaluated the Spiking-Hybrid method on two sets of in-situ data: 299 sorghum samples grown in the greenhouse and 566 maize samples grown in the field. Leaf VIS-NIR-SWIR reflectance data (400–2500 nm) were measured concurrently with Cab and Cw sampling. We used a leaf-level radiative transfer model (PROCOSINE) to simulate leaf reflectance as the function of leaf biochemical and internal structural parameters. The machine learning models included partial least squares regression, Gaussian process regression, and gradient boosting regression. The proposed method was further compared with three other common methods in plant trait estimation: the PROCOSINE inversion with numerical optimization, the conventional hybrid method, and the empirical machine learning method. Results showed that the proposed and the empirical methods outperformed the other two methods for estimating Cab and Cw, indicating the importance of involving in-situ data in model training. Moreover, we explored the effect of in-situ sample size. When the number of available in-situ samples was limited (i.e., less than 5 % of the total training samples), the Spiking-Hybrid method obtained consistently higher accuracy and better robustness than the empirical machine learning method. These results imply that the Spiking-Hybrid method can achieve accurate and robust estimations for Cab and Cw by harnessing the free simulated data and a significantly reduced number of in-situ samples. This method has great potential for agricultural research with limited in-situ samples.