Estimation of Annual Productivity of Sown Rainfed Grasslands Using Machine Learning

Abstract

Grasslands play a critical role in providing diverse ecosystem services. Sown biodiverse pastures (SBP) rich in legumes are an important agricultural innovation that increases grassland productivity and reduces the need for fertilisers. This study developed a machine learning model to obtain spatially explicit estimations of the productivity of SBP, based on field sampling data from five Portuguese farms during four production years (2018–2021) and under two fertilisation regimes (conventional and variable rate). Weather data (such as temperature, precipitation and radiation), soil properties (including sand, silt, clay and pH), terrain characteristics (including elevation, slope, aspect, hillshade and topographic position index), and management data (including fertiliser application) were used as predictors. A variance inflation factor (VIF) approach was used to measure multicollinearity between input variables, leading to only 11 of the 53 input variables being used. Artificial neural network (ANN) methods were used to estimate pasture productivity, and hyper-parameterization optimization was performed to fine-tune the model. Plots under variable rate fertilisation were significantly improved by up to 20 kg P ha⁻¹ applied in the same year. Plots under conventional fertilisation benefitted the most from fertilisation in past years. The model demonstrated good generalisation, with similar estimation errors for both the training and test sets: for an average yield of 6096 kg ha⁻¹ in the sample, the root mean squared errors (RMSE) for the training and test sets were respectively 882 and 1125 kg ha⁻¹. These results indicate that the model did not overfit the training data and can be used to estimate SBP productivity maps in the sampled farms. However, further studies are required to asses if the obtained model can be applied to new unseen data.