Prior knowledge and active learning enable hybrid method for estimating leaf chlorophyll content from multi-scale canopy reflectance

Abstract

Real-time and accurate assessment of leaf chlorophyll content (Cab) will be significant for monitoring plant physiological status. Development of hybrid methods advances the assessment of Cab with the advantages of both physical mechanism and calculation capability, while the estimation accuracies were constrained by the lacking of prior knowledge and representative real samples. This study explored the use of prior knowledge and active learning to enable the development of efficient hybrid methods for assessing Cab from multi-scale canopy reflectance. The hybrid method was established by using the simulated dataset from the PROSAIL model to train the Gaussian process regression (GPR) model. We used two measured crop datasets and three publicly-available grassland, herbaceous, and tree species datasets collected from near-ground, unmanned aerial vehicle, and airborne platforms to examine the model performances. Prior knowledge for model parameters was obtained from ground measurements, and an active learning method, namely improved Greedy sampling (iGS), was employed to select new representative samples. The results showed that performance of the hybrid method in Cab estimations varied with leaf structure parameter, carotenoid content, anthocyanin content, carbon-based constituents, average leaf angle, leaf area index, and hot-spot size parameter. When prior knowledge for the ranges of model parameters was available, the hybrid method better estimated Cab with the root mean square error (RMSE) reduced by 7.59%–65.96%. Compared with the random selection, iGS was able to select representative samples to update the estimation models with only 10% of the measured samples, and a higher ratio (e.g., 20%) may be needed for small datasets. The hybrid method coupled with prior knowledge and active learning obtained satisfied estimations of Cab across different datasets with the relative RMSE of 12.30%–27.24%. The proposed method will contribute to improving the applicability and robustness of hybrid modeling for multi-scale monitoring of leaf biochemical traits.