Feature engineering on climate data with machine learning to understand time-lagging effects in pasture yield prediction

Abstract

Pastures are a primary food source for livestock in Australia, with various pasture species grown in rotations. Accurate prediction of pasture availability is critical for effective farm management, livestock growth, and maintaining the supply chain. Environmental factors, particularly climate, heavily influence pasture yield. However, different pasture species respond to climate attributes with varying time lags; for example, one species might be more influenced by last week’s weather while another by the previous month’s highlighting the nuanced temporal dependencies. This time-lagging effect complicates the development of machine-learning models that can learn the temporal dependencies to predict pasture yield. To address this, our study proposes an averaging-based feature engineering approach, effectively capturing the varying temporal dependencies across pasture species and also allowing interpretation of the dependencies. Utilizing remote sensing and climate data, covering 196 farms (and 6885 paddocks) across Australia, we applied several machine learning techniques, including XGBoost, random forest, linear regression, deep neural networks, stacking, and bootstrapping. Our results show that incorporating averaging-based feature-engineered climate attributes significantly improves pasture yield predictions, with enhancements of up to 20.28%, 31.81%, and 31.11% across the three evaluation measures, RMSE, MAE, and R${}^2$, respectively. This approach also enhances interpretability, revealing diverse time-lagging effects on different pasture species. XGBoost-based feature importance analysis further unveils insights into the influence of each climate attribute and its temporal dependencies on pasture yield.