Informing the prediction of forage quality of Mediterranean grasslands using hyperspectral reflectance: Concentration vs content, phenology, and generalisation of models

Context

Remote sensing has shown potential to provide accurate and real-time information on grassland forage quality, crucial for the management of livestock systems. However, there are still uncertainties that must be considered to make models reliable and practical. A source of discrepancy regards the measurement unit describing forage quality variables, namely either mass-based concentration (%) or mass per surface area content (kg ha⁻¹). Furthermore, phenological patterns largely influence grassland reflectance and have a strong impact on model accuracy. Also, the generalisation of empirical models in heterogeneous grasslands can hinder their applicability.

Objectives

1) Assess the suitability of retrieving forage quality parameters as concentration (%) versus content (kg ha⁻¹). 2) Investigate the performance of multitemporal compared to phenophase-specific models. 3) Evaluate the generalisation ability of the models.

Methods

Samples were collected from five farms to determine Dry Matter Yield (DMY) and both, concentration (%) and content (kg ha⁻¹) of forage quality variables including crude protein (CP), neutral and acid detergent fibre (NDF, ADF), and enzyme digestibility of organic matter (EDOM). The relationship between forage quality variables and DMY were analysed by Pearson Correlations and Principal Component Analysis. Reflectance was recorded with a FieldSpec spectroradiometer. Partial Least Squares Regression (PLSR) was used to explore the relationship between forage variables and reflectance.

Results

The forage quality variables were strongly correlated to DMY when expressed as content (kg ha⁻¹) r>0.83 but not when expressed as concentration (%). For the best predicted variables, CP and NDF, the results of the PLSR models indicated better performance in concentration-based estimation. CP% was the best predicted variable (R²_cv=0.8, NRMSE_cv=9.8 %). Multitemporal models showed overall higher performance (CP%, R²_cv=0.81) than phenophase-specific models (CP%, R²_cv=0.60 green, R²_cv=0.70 green-senescent and R²_cv<0 senescent grasslands). The generalisation ability was low and varied among farms (R²_test 0–0.60).

Conclusions

The use of concentration (%) is more accurate and representative of forage quality than content (kg ha⁻¹), which seemed redundant with DMY and misleading from the true nutritive value for livestock.

Multitemporal models performed better than phenophase-specific models due to their larger range of values. The ability to predict forage quality in senescent grasslands is low. The usefulness of the models is context-dependent, and their application requires knowledge of the limitations and status of the grasslands. Efforts must be directed toward improving the generalisation ability through the development of models calibrated with larger and more diverse datasets.