Abstract
Yield, crude protein content and crude protein yield are relevant parameters for grassland farmers. The objective of this study is to assess the potential of proximal sensing as an alternative to traditional methods (forage sampling and NIRS-analysis), for predicting these parameters. Specifically, predictive models were developed based on (i) hyperspectral measurements from a field spectroradiometer and (ii) the same data resampled to Sentinel-2 satellite bands. Measurements were conducted on 72 plots during the 2nd and 3rd harvests of a multispecies experiment carried out within the framework of the LegacyNet project, investigating mixtures of six species from three functional groups arranged as a simplex design and four more complex seed mixtures for leys. In each plot, hyperspectral measurements were coupled to results from laboratory forage quality analysis. Data preprocessing included Multiplicative Scatter Correction, Savitzky–Golay smoothing, and derivatives. Partial Least Squares Regression models were trained on 70% of the data and evaluated on 30% using R 2 and RMSE as metrics. Dry matter yield was predicted with R 2 values between 0.78 and 0.89, crude protein content between 0.72 and 0.83, and crude protein yield between 0.64 and 0.74. Despite the limited sample size, results for predicting these parameters in multispecies swards are promising.