Machine learning estimators for the quantity and quality of grass swards used for silage production using drone-based imaging spectrometry and photogrammetry

Abstract Drones offer entirely new prospects for precision agriculture. This study investigates the utilisation of drone remote sensing for managing and monitoring silage grass swards. In northern countries, grass swards are fertilised and harvested three times per season when aiming to maximise the yield. Information about the grass quantity and quality is necessary to optimise these operations. Our objectives were to investigate and develop machine-learning techniques for estimating these parameters using drone photogrammetry and spectral imaging. Trial sites were established in southern Finland for the primary growth and regrowth of grass in the summer of 2017. Remote-sensing datasets were captured four times during the primary growth season and three times during the regrowth period. Reference measurements included fresh and dry biomass and several quality parameters, such as the digestibility of organic matter in dry matter (the D-value), neutral detergent fibre (NDF), indigestible neutral detergent fibre (iNDF), water-soluble carbohydrates (WSC), the nitrogen concentration (Ncont) in dry matter (DM) and nitrogen uptake (NU). Machine-learning estimators based on random forest (RF) and multiple linear regression (MLR) methods were trained using the reference measurements and tested using independent test datasets. The best results for the biomass estimation, nitrogen amount and digestibility were obtained when using hyperspectral and 3D data, followed by the combination of multispectral and 3D data. During the training process, the best normalised root-mean-square errors (RMSE%) were 14.66% for the dry biomass and 12% for fresh biomass; the best RMSE% values for NU, the D-value and NDF were 13.6%, 1.98% and 3% respectively. For the primary growth, the accuracies of all quality parameters were better than 20% with the independent test datasets; for the regrowth, the estimation accuracies of the D-value, iNDF, NDF, Ncont and NU were better than 20%. The results showed that drone remote sensing was an excellent tool for the efficient and accurate management of silage production.