Use of different machine learning algorithms and their ability to predict organic matter digestibility in grass from constituent or spectral data

Authors: OLLERTZ-MERTENS, BERND - Norwegian University Of Life Sciences; VOLDEN, HARALD - Norwegian University Of Life Sciences; PRESTLOKKEN, EGIL - Norwegian University Of Life Sciences; DIGMAN, MATTHEW - University Of Wisconsin; Kalscheur, Kenneth; BUCHNER, BENJAMIN - John Deere & Company

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/9/2026
Publication Date: 1/10/2026
Citation: Ollertz-Mertens, B., Volden, H., Prestlokken, E., Digman, M.F., Kalscheur, K., Buchner, B. 2026. Use of different machine learning algorithms and their ability to predict organic matter digestibility in grass from constituent or spectral data. Smart Agricultural Technology. Volume 13, 101792. https://doi.org/10.1016/j.atech.2026.101792.
DOI: https://doi.org/10.1016/j.atech.2026.101792

Interpretive Summary: Grassland accounts for roughly two-thirds of agricultural land worldwide, and is a critical resource required for ruminant livestock production. Organic matter digestibility at harvest is the main factor determining nutritive values of grass silage. Organic matter digestibility declines rapidly throughout the growing season, reducing the availability of energy for production of milk. To improve the prediction and forecasting of organic matter digestibility, the use of machine learning can be implemented to predict the quality of grass. The objective of this research was to determine which machine-learning approach is best suited to forecast organic matter digestibility in grass using near-infrared spectroscopy technology as input. Results concluded that near-infrared spectroscopy technology is able to predict organic matter digestibility changes in field in real-time, allowing farmers to adjust their management accordingly to harvest higher quality forages. This research will be of interest to livestock producers, agronomists, forage scientists, animal scientists, animal nutritionists, and livestock system researchers interested in how machine learning approaches can be used to forecast nutritional quality of forages. This may be a useful tool in forage harvesting and feeding strategies to improve performance of dairy cows fed high forage diets.

Technical Abstract: Organic matter digestibility (OMD) at harvest is the main factor determining nutritive values of grass silage. The OMD declines rapidly throughout the growing season, reducing availability of energy for production of milk and meat, potentially decreasing feed efficiency and increasing methane yield. Thus, information on OMD is critical to determining when to harvest grass for silage. The objective of this study was to identify machine learning (ML) algorithms best suited to predict OMD from field-acquired near-infrared spectroscopy (NIRS) data. Fourteen different ML models were trained on constituent predictions from NIRS devices. Furthermore, eleven models were trained on spectral data from said devices. Open-source weather data was also included in those models. Model performance was evaluated based on the coefficient of determination (R²) and the root mean squared error (RMSE). Samples from first to fourth cuts were collected from seventeen fields, on six different farms, in three different years. Ten forage constituents were predicted using NIRS, moreover spectral data were recorded. Laboratory-based NIRS predictions served as reference data. Using constituent data as an input, seven models reached the target of R² above 0.7, but none reached the target of RMSE =3. Gradient boosting regression performed best of all considered models in this application (R² ˜ 0.82-0.86; RMSE ˜ 3% of dry matter). Using raw spectral data did not improve performance compared to NIRS-predicted constituent data. However, reducing dimensions using principal component regression resulted in the best-performing model overall, showing an RMSE of 2.76% of dry matter (DM). Based on the results, it is concluded that NIRS technology is able to predict OMD changes in-field in real-time, allowing farmers to adjust their management accordingly. Continued work on those models to improve their prediction accuracy and widen their usability with respect to areas, swards, and farm setups is proposed.