# Non-destructive estimation of maize carotenoids using reflectance-based spectral indices

## TL;DR

Researchers developed new methods to non-destructively estimate carotenoid levels in maize leaves using spectral reflectance and machine learning.

## Contribution

New spectral indices (CAR7, CAR8, CAR9) and a machine learning model (REPTree) were developed for improved non-destructive carotenoid estimation in maize.

## Key findings

- Carotenoid concentration significantly affects leaf reflectance in the visible range (500–650 nm).
- New spectral indices CAR7, CAR8, and CAR9 showed strong predictive performance (R² = 0.72–0.83).
- REPTree machine learning model provided the most reliable carotenoid estimation (R² = 0.79, NRMSE = 13.84%).

## Abstract

This study investigates the relationship between maize leaf carotenoid content and spectral reflectance, evaluates existing carotenoid estimation indices, and develops new spectral indices and machine learning models for improved prediction. A strong positive correlation was observed between carotenoid and chlorophyll content, highlighting carotenoids’ role in both light harvesting and photoprotection. Spectral analysis revealed that carotenoid concentration significantly affects leaf reflectance in the visible range, particularly between 500–650 nm. Existing carotenoid indices exhibited limited predictive performance for the studied samples, prompting the development of nine new indices based on principal component analysis. Among these, CAR7, CAR8, and CAR9 demonstrated superior predictive ability across different training (2021-2022: R2 = 0.72-0.76, NRMSE = 15-16%, 2021-2023: R2 = 0.60-0.62, NRMSE = 11-12%, 2022-2023: R2 = 0.42-0.49, NRMSE = 18.3-18.5%) and testing periods (2023: R2 = 0.44-0.50, NRMSE = 14-19%, 2022: R2 = 0.65-0.72, NRMSE = 13-16%, 2021: R2 = 0.81-0.83, NRMSE = 18.28-24.65%). Machine learning models further improved carotenoid estimation, with REPTree providing the most reliable and balanced performance during testing (R2 = 0.79, NRMSE = 13.84%). The findings suggest that the combination of targeted spectral indices and appropriate machine learning approaches enables accurate, non-destructive estimation of maize carotenoid content, offering potential for practical applications in crop monitoring and stress assessment.

The abstract should ideally be structured according to the IMRaD format (Introduction, Methods, Results and Discussion). Provide a structured abstract if possible. If your article has been copyedited by us, please provide the updated abstract based on this version.Flowchart detailing a study process. It includes sections on study design, leaf sampling, laboratory analysis, spectral measurements, data processing, feature selection, development of spectral indices, model calibration and validation, machine learning modeling, and model evaluation. Each section lists specific methodologies and parameters such as sampling times, spectral calibration, data integration, use of PCA, machine learning techniques, and evaluation metrics. Key performers are CAR7, CAR8, CAR9, and REPTree. The study site is Hungary, from 2021 to 2023, with both irrigated and non-irrigated plots.

The abstract should ideally be structured according to the IMRaD format (Introduction, Methods, Results and Discussion). Provide a structured abstract if possible. If your article has been copyedited by us, please provide the updated abstract based on this version.