Hyperspectral Leaf Reflectance as Proxy for Photosynthetic Capacities: An Ensemble Approach Based on Multiple Machine Learning Algorithms

Abstract

Global agriculture production is challenged by increasing demands from rising population and a changing climate, which may be alleviated through development of genetically improved crop cultivars. Research into increasing photosynthetic energy conversion efficiency has proposed many strategies to improve production but have yet to yield real-world solutions, largely because of a phenotyping bottleneck. Partial least squares regression (PLSR) is a statistical technique that is increasingly used to relate hyperspectral reflectance to key photosynthetic capacities associated with carbon uptake (maximum carboxylation rate of Rubisco, V c,max ) and conversion of light energy (maximum electron transport rate supporting RuBP regeneration, J max ) to alleviate this bottleneck. However, its performance varies significantly across different plant species, regions, and growth environments. Thus, to cope with the heterogeneous performances of PLSR, this study aims to develop a new approach to estimate photosynthetic capacities. A framework was developed that combines six machine learning algorithms, including artificial neural network (ANN), support vector machine (SVM), least absolute shrinkage and selection operator (LASSO), random forest (RF), Gaussian process (GP), and PLSR to optimize high-throughput analysis of the two photosynthetic variables. Six tobacco genotypes, including both transgenic and wild-type lines, with a range of photosynthetic capacities were used to test the framework. Leaf reflectance spectra were measured from 400 to 2500 nm using a high-spectral-resolution spectroradiometer. Corresponding photosynthesis vs. intercellular CO 2 concentration response curves were measured for each leaf using a leaf gas-exchange system. Results suggested that the mean R 2 value of the six regression techniques for predicting V c,max ( J max ) ranged from 0.60 (0.45) to 0.65 (0.56) with the mean RMSE value varying from 47.1 (40.1) to 54.0 (44.7) μmol m -2 s -1 . Regression stacking for V c,max ( J max ) performed better than the individual regression techniques with increases in R 2 of 0.1 (0.08) and decreases in RMSE by 4.1 (6.6) μmol m -2 s -1 , equal to 8% (15%) reduction in RMSE . Better predictive performance of the regression stacking is likely attributed to the varying coefficients (or weights) in the level-2 model (the LASSO model) and the diverse ability of each individual regression technique to utilize spectral information for the best modeling performance. Further refinements can be made to apply this stacked regression t...

Estimating photosynthetic traits from reflectance spectra: A synthesis of spectral indices, numerical inversion, and partial least square regression

“…This study provided direct evidence that mapping V cmax and J max at the canopy level could be successful with reflectance spectra (or derived variables) from 400 to 900 nm used as predictors. The findings shown in Figure were in agreement with previous studies to show the PLSR model as an effective tool to predict photosynthetic variables across different spatial scales (Ainsworth et al, ; Dechant et al, ; Fu et al, ; Meacham‐Hensold et al, ; Serbin et al, ; Serbin et al, ). However, this study only used reflectance spectra from 400 to 900 nm for the PLSR modelling, which exhibited an even higher or at least similar R 2 value compared with previous modelling results at the leaf level (e.g., Dechant et al, ).…”

“…Results presented in this study showed that V cmax and J max could be well estimated for 11 cultivars of one crop (including both genetically modified and wild types) using the proposed three approaches (i.e., PLSR of reflectance spectra, spectral indices, and RTM‐inversion of crop traits) with R 2 all larger than 0.6. These modelling performances at the canopy level using reflectance spectra from 400 to 900 nm were even better than those at the leaf level using a similar dataset from the whole spectrum (400–2,500 nm; Fu et al, ; Meacham‐Hensold et al, ). A similar finding was also observed in comparisons between Serbin et al () and Serbin et al () for tree species: V cmax were better predicted at the canopy level ( R 2 = 0.94) than at the leaf level ( R 2 = 0.89).…”

“…However, as the variance in V cmax and J max (as shown in Figure b) for tobacco cultivars is generally larger than that for food crops such as maize and soybean, research efforts are still needed to evaluate the developed approaches in high‐throughput phenotyping of food crops for desired photosynthetic improvements. More importantly, there is a need to explore whether the developed approaches for estimating photosynthetic capacities are species dependent (Fu et al, ).…”

“…In this study, a total of 48 data pairs were collected for V cmax and 39 data pairs for J max at the plot level. The fewer number of measurements for V cmax than that for J max resulted from one of the genetically modified tobacco cultivars not being electron transport limited under any conditions (removed from analysis as suggested in Fu et al ).…”

“…Given these reflectance‐leaf characteristics associations, machine learning algorithms such as partial least square regression (PLSR; Geladi & Kowalski, ; Wold et al, ) and artificial neural network regression (Specht, ) are widely used for analysing hyperspectral data because of their ability to deal with irrelevant spectral bands and band collinearity, also known as the curse of dimensionality of data (Thenkabail et al, ). The understanding of the physiological mechanism for correlating reflectance spectra with photosynthetic variables, however, remained unsolved using complex machine learning algorithms (Fu et al, ). Alternatively, vegetation indices (VIs) such as normalized difference VI, photochemical reflectance index, and chlorophyll index have also been used to reveal photosynthetic productivity (Ainsworth et al, ; Drolet et al, ; DuBois et al, ; Gamon et al, ; Muraoka et al, ).…”