| Abstract | The potential of grasslands' fodder production is a crucial management measure, while its quantification is still laborious and costly. Remote sensing technologies, such as hyperspectral field measurements, enable fast and non-destructive estimation. However, such methods are still limited in transferability to other locations or climatic conditions. With this study, we aim to predict forage nutritive value, quantity, and energy yield from hyperspectral canopy reflections of grasslands across three climate zones. We took hyperspectral measurements with a field spectrometer from grassland canopies in temperate, tropical and semi-arid grasslands, and analyzed corresponding biomass samples for their quantity (BM), metabolizable energy content (ME) and metabolizable energy yield (MEY). Three machine learning algorithms were used to establish prediction models for single and across climate regions. Tue normalized root mean squared error (nRMSE) for ME, BM and MEY varied between 0.12 - 0.19, 0.14 - 0.21, and 0.15 - 0.21, respectively. Tue ME trans-climatic model showed the best accuracy compared to the local models. Trans-climatic model predictions of climate-specific data, decrease in accuracy to 0.16- 0.21, 0.17 - 0.24, and 0.19- 0.28 for ME, BM and MEY compared to predictions with climate-specific models. Trans-climatic models with feed-forward neural networks showed similar performance for ME but higher accuracies for BM and MEY predictions. Tue trans-climatic models generally showed good performance for forage nutritive value and forage provision. Our results suggest that models based on hyperspectral measurements offer great potential to assess or even map the forage nutritive value of grasslands across climate zones. |
|---|
