Hyperspectral Reflection and Forage Data on Grassland Vegetation

With the following dataset we aim to predict forage quantity and quality proxies—namely aboveground biomass (AGB), metabolizable energy content (ME), and metabolizable energy yield (MEY)—using hyperspectral canopy reflectance data from grassland communities across diverse climatic zones. We hypothesize that spectral data will facilitate reliable predictions of forage traits across grassland biomes.

Data Collection
Data were collected from three distinct grassland ecosystems: a seasonally humid tropical savanna in West Africa (“Tropical”), a dry subtropical thornbush savanna in Namibia (“Semi-arid”), and temperate grasslands in Germany (“Temperate”). We gathered 456 hyperspectral measurements (177 from Germany in 2020-2021, 105 from West Africa in 2012, 174 from Namibia in 2021) using an ASD FieldSpec 3 in Germany and West Africa, and an ASD FieldSpec 4 in Namibia. Reflectance spectra were recorded from 350 to 2,500 nm at 1.5 m (Namibia/Germany) and 1.3 m (West Africa) above ground, calibrated with Spectralon® standards. Afterwards, aboveground biomass was clipped, with dry weight determined after oven-drying (55–60°C >48 hours). ME was assessed using in vitro gas production after Menke & Steingass, and crude protein content was measured via the Kjeldahl method or an elemental analyser. MEY was calculated as the product of ME and AGB.

Data Processing
Spectral data processing involved averaging five replicate measurements and applying a Savitzky-Golay filter for smoothing. Spectral regions affected by noise and atmospheric absorption were excluded. Data pre-processing was conducted in R using the “hsdar” package, resulting in a predictor set derived from reflectance spectra reduced to 10 nm resolution, yielding 168 predictors. The first derivative provided an additional 153 predictors, and vegetation indices contributed another 170, totaling 491 predictors. Each datapoint of ME, AGB, and MEY was assigned the full predictor set, available in “Pred_ME.csv”, “Pred_BM.csv”, and “Pred_MEY.csv”.

Model Setup
Prediction models were established using Random Forest (RF), Partial Least Squares (PLS) Regression Algorithms and convolutional neural networks (cNN). Predictor reduction was executed through repeated backward elimination based on predictor importance. Optimal predictor sets for both models were determined by minimizing mean square error (MSE) via repeated 10-fold cross-validation. The first derivative predictor set was finally used for RF, PLS and cNN model setup.

Data availability and usage
The predictor reduction and model setup codes for PLS and RF are available in “ModelSetup_TransClim_RF-PLS.R”, while the cNN model setup code is in “ModelSetup_TransClim_DL.py”. The input datasets can be used for model training with the 491 possible predictors and ME, AGB/BM and MEY as response variables. The codes are applicable to the ME-Dataset and can be adapted to AGB (BM) and MEY accordingliy if used in one project folder.