Multispectral imaging for zeaxanthin content in the exocarp of chili peppers.

Abstract

This study developed a model to predict zeaxanthin content in peppers using multispectral imaging and chemical data. A one-dimensional convolutional neural network (1D CNN) model was identified as the optimal single-modal model after comparing four machine learning algorithms. On the prediction dataset, the model achieved a determination coefficient ( Rp 2 ) of 0.7639. Building upon the 1D CNN framework, a multimodal feature fusion model (MCSF) was constructed by integrating the chemical measurements of capsanthin and total carotenoid contents using a multilayer perceptron. This enhanced model demonstrated excellent predictive accuracy and robustness, with Rp 2 values of 0.9318 and 0.9211 across different spectral ranges. For high-throughput detection purposes, a simplified model that replaced measured capsanthin with a comprehensive red index still performed well, with an Rp 2 of 0.8912 and an RPD of 3.11. This strategy provides a new solution for the efficient spectral detection of plant chemicals affected by multicollinearity in their absorption spectra.

Plant phenotyping relevance

マルチスペクトル画像と機械学習を用いて、トウガラシ果皮のゼアキサンチン含量という植物器官の形質を非破壊・高スループット推定する手法を開発・評価しており、フェノタイピング手法が中心である。

Code and data availability

The paper's data availability statement explicitly states that the datasets (multispectral imaging and chemical trait measurements) and the main model code are publicly available in the authors' GitHub repository, which is an allowed URL.