Causal forest approach for site-specific input management via on-farm precision experimentation

Abstract

Estimating site-specific crop yield response to changes to input (e.g., seed, fertilizer) management is a critical step in making economically optimal site-specific input management recommendations. Past studies have attempted to estimate yield response functions using various Machine Learning (ML) methods, including the Random Forest (RF), Boosted Random Forest (BRF), and Convolutional Neural Network (CNN) methods. This study proposes use of the Causal Forest (CF) model, which is one of the emerging ML methods that comprise “Causal Machine Learning.” Unlike previous yield-prediction-oriented ML methods, CF focuses strictly on estimating heterogeneous treatment effects (changes in yields that result from changes in input application rates) of inputs. We report results of using Monte Carlo simulations assuming various production scenarios to test the effectiveness of CF in estimating site-specific economically optimal nitrogen rates (EONRs), comparing CF with the yield-prediction-oriented ML methods RF, BRF, and CNN. CF’s estimations of site-specific EONRs were superior under all scenarios considered. We also show that the model’s yield prediction accuracy need not imply EONR prediction accuracy.