Estimation of botanical composition of forage crops using laboratory-based hyperspectral imaging and near-infrared spectrometer measurements

• Published in: Journal of Agriculture and Food Research
• Main Authors: PENG Junxiang, RAHIMI JAHANGIRLOU Maryam, MOREL Julien, ZHOU Zhenjiang, PARSONS David
• Language: English
• Published: Elsevier BV 2024
• Subjects: Nash; Sutcliffe; Northern Sweden
• Online Access: https://publications.jrc.ec.europa.eu/repository/handle/JRC137962; https://www.sciencedirect.com/science/article/pii/S2666154324003569?dgcid=rss_sd_all; https://doi.org/10.1016/j.jafr.2024.101319

Harvested forage is the main raw feed for ruminant animals in Sweden, and is commonly cultivated in mixed stands of legume and grass species. The fraction of legume on a dry matter basis, known as botanical composition (BC) is a very important indicator of forage quality. In this study, hyperspectral imaging and near-infrared spectrometer (NIRS) based methods were used to estimate BC, to overcome the shortcomings of hand separation, which is time and resource consuming. Samples of mixes of timothy and red clover were collected from 2017 to 2019 from multiple sites in Northern Sweden and hand separated. The samples were synthetically mixed to 11 different BC levels, i.e., 0-100% clover content. Two different instruments (Specim SWIR hyperspectral imaging system and Foss 6500 spectrometer) were used to collect spectral data of samples milled to two levels of coarseness. Three different regression analyses: partial least square regression (PLSR), support vector regression (SVR) and random forest regression (RFR), were used to build BC estimation models. The effects of the milling particle sizes and the different sensors on the performances of the models were compared. Among different regression analyses, PLSR performed best, with Nash-Sutcliffe efficiency (NSE) for model evaluation from 0.79 to 0.89, varying for different sensors and milling sizes, and corresponding root mean square error (RMSE) from 10.29% to 14.4%. SVM had evaluation NSE from 0.56 to 0.81, and corresponding RMSE from 13.85% to 21.06%. RFR had evaluation NSE from -0.15 to 0.39, and corresponding RMSE from 24.71% to 33.84%. Finer milling made the model accuracies slightly higher. The Specim camera showed slightly better performances than the Foss spectrometer with finely milled samples, with NSE of 0.89 and RMSE of 10.3% for the PLSR model evaluation. This study developed quick and robust methods to determine the BC of timothy grass and red clover mixtures, which can provide useful information for farmers or researchers. JRC.D.5 - Food Security