Evaluation of forage grass biomass estimation models using multispectral drone imaging across multiple sites

Abstract. Sustainable grassland management practices enhance key ecological functions, including carbon sequestration, biodiversity conservation, and the maintenance of soil fertility essential for climate change mitigation. Accurate and reliable estimation of grass biomass is essential for decision making on harvesting time and rate of fertilizer application. Remote sensing and data analysis technologies offer unprecedented opportunities for monitoring grassland dynamics, yet methodological challenges persist in generalizing remote sensing-based models for different growths and different areas. This study investigates the estimation of grass biomass of different growth stages during two years using multispectral UAS-based remote sensing. A leave-one-out cross-validation was conducted using five harvest datasets to train and test random forest (RF) and partial least squares regression (PLSR) models, assessing estimation accuracy within individual sites. This was followed by a cross-site evaluation, where models trained using data from other locations were tested on each harvest date to evaluate model generalizability. The estimation models within the Maaninka site yielded at the best NRMSEs 14.7%, but exceeded 55% on two cutting dates. Incorporating data from multiple sites improved generalization or maintained similar accuracy across test dates. The findings indicated that using data from various locations can improve model stability, especially in cases where local data does not provide strong predictive information.