ced by the quality and size of the datasets used for training the model. To minimize the “over-fitting” error, the large dataset was used and split carefully to include the different multi-cereal species in both the calibration and validation dataset (as described in section 2.5.1). The code is available on the GitHub platform (https://github.com/adamavip/nirs-protein-prediction) and its particular parts can be now utilized to enhance and develop other pipelines and products. For our dataset, the algorithms built using ML-based methods (particularly, the stacked ensemble model via Hone Create and custom-designed CNN; section 3.4) achieved the better comparative metrics for both spectra types
Open resource ↗adamavip/nirs-protein-prediction · pdf-raw-page:14 lines:1-54Paper record
Machine Learning-Powered Models for Near-Infrared Spectrometers: Prediction of Protein in Multiple Grain Cereals
25 Feb 2022 · 10.20944/preprints202202.0335.v1
Abstract
Achieving global goals on sustainable nutrition, health, and wellbeing will depend on delivering enhanced diets to humankind. This will require, among others, instantaneous access to information on food quality at key points within agri-food systems. Although stationary methods are usually used to quantify grain quality (wet-lab chemistry, benchtop NIR spectrometer); these do not suit many required user-cases, such as stakeholders in decentralized agri-food-chains that are typical for emerging economies. Therefore, we explored new technologies and models that might aid these particular user-cases. For this purpose, we generated the NIR spectra of 328 grain samples from multiple cereals (finger millet, foxtail millet, maize, pearl millet, sorghum) with a standard benchtop NIR Spectrometer (DS2500, FOSS) and a novel mobile NIR-based sensor (HL-EVT5, Hone). We explored a range of classical deterministic and novel machine learning (ML)-driven models to build calibrations out of the NIR spectra. We were able to build relevant calibrations out of both types of spectra. At the same time, ML-based methods enhanced the prediction capacity of calibration models compared to classical deterministic methods. We also documented that the prediction of grain protein content based on NIR spectra generated by a mobile sensor (HL-EVT5, Hone) was highly relevant for quantitative protein predictions (R2 = 0.91, RMSE = 0.97, RPD = 3.48). Thus, the findings of this study lay the foundations on which to expand the utilization of NIR spectroscopy applications for agricultural research and development.
Code and data availability