Bayesian calibration of management practices for a crop model implemented in a subsistence farming region

Abstract

Rainfed agriculture is crucial for food security in sub-Saharan Africa, yet it faces significant challenges from climate variability, soil degradation, and limited access to resources. Process-based crop models are widely used in agricultural research as well as in decision support systems. These systems play an important role in aiding policymakers in designing and implementing strategies to enhance food security. Farm management practices are one essential input for crop models. However, that data exhibit farm-scale variabilities and is usually scarce in regions with fragile food production systems, rendering the powerful crop modeling tools ineffective, particularly in large-scale applications. We present a new approach to infer the relevant management practices of a region in a data scarce environment. We introduce Bayesian calibration as a method to infer key management practices using the CERES-Maize model within DSSAT, in order to provide more reliable yield estimates in a subsistence-farming region. This novel approach allows to better represent the uncertainty in the unknown input management practices in addition to the soil and weather-related variabilities. A study case was presented using farm-level maize yield data from 18 wards in North-western Zimbabwe from the 2021/22 season. The calibrated model provided reliable yield estimates for 72 % of the wards, significantly outperforming the non-calibrated model, which captured the observed yield for only 22 % of the wards. Furthermore, the calibrated model better captured intra-regional yield variation, with an R² of 0.42 and a d-agreement index of 0.67. This approach underscores the importance of accurately representing the variability of management practices in larger-scale implementations of crop models. This approach will allow the crop models to be effectively used for monitoring and forecasting of crop yield for a wide swath of fragile lands with limited data availabilities.