Predicting pasture biomass using a statistical model and machine learning algorithm implemented with remotely sensed imagery

• Published in: Computers and electronics in agriculture Vol. 180; p. 105880
• Main Authors: De Rosa, Daniele, Basso, Bruno, Fasiolo, Matteo, Friedl, Johannes, Fulkerson, Bill, Grace, Peter R., Rowlings, David W.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01-01-2021; Elsevier BV
• Subjects: Algorithms; Biomass; Cameras; Environmental management; Generalised additive model; Grazing; Image acquisition; Machine learning; Multispectral imagery; Pasture; Profitability; Random forest; Remote sensing; Spatial resolution; Statistical models; UAV; Unmanned aerial vehicles; Multispectral imagery; Random forest; UAV; Pasture; Generalised additive model; Remote sensing
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105880

•GAM and ML models can be used as a farm-based decision support tool.•Decision support tools can be used in combination with remotely sensed imagery.•High spatio-temporal variability of pasture biomass was observed in pasture fields.
Accurate daily estimates of pasture biomass can improve the profitability of pasture-based dairy system by optimising input of feed supplements and pasture utilisation. However, obtaining accurate pasture mass estimates is a laborious and time-consuming task. The aim of this study was to test the performance of an integrated method combining remote sensing imagery acquired with a multispectral camera mounted on an unmanned aerial vehicle (UAV), statistical models (generalised additive model, GAM) and machine learning algorithms (random forest, RF) implemented with publicly available data to predict future pasture biomass loads. This study showed that using observations of pasture growth along with environmental and pasture management variables enabled both models, GAM and RF to predict the pre-grazing pasture biomass production at field scale with an average error below 20%. If predictive variables (i.e. post-grazing pasture biomass) were excluded, model performance was reduced, generating errors up to 40%. The post-grazing biomass information at high spatial resolution (<1 m) acquired with the UAV-multispectral camera system was used as predictive variable for future pasture biomass. With the inclusion of the spatially explicit post-grazing biomass variable both models accurately predicted the pre-grazing pasture biomass with an error of 27.7% and 22.9% for RF and GAM, respectively. However, the GAM model performed better than RF in reproducing the spatial variability of pre-grazing pasture biomass. This study demonstrates the capability of statistical and machine learning models implemented with UAV or manually obtained pasture information along with publicly available data to accurately predict future pasture biomass at field and farm scale.