Investigating the impact of coupling process-based and data-driven models on wheat crops in arid and semi-arid regions
Date published
Free to read from
Authors
Supervisor/s
Journal Title
Journal ISSN
Volume Title
Publisher
Department
Type
ISSN
Format
Citation
Abstract
Accurate prediction of wheat yields in arid and semi-arid regions is challenging due to water scarcity, varying environmental conditions, and the dynamic nature of factors influencing crop growth. This thesis aims to enhance scalable wheat yield prediction by integrating remote sensing (RS) data into process-based and data-driven models for more precise and accurate yield prediction in these regions, supporting both tactical and strategic decision-making in agriculture. AquaCrop was chosen for its robust simulation of crop yield response to water. Four interlinked research questions are addressed in this study. First, I identify key factors impacting wheat yield prediction based on sensitivity and SHAP analysis for process-based and data-driven models, respectively. Second, I compare the trade-offs between calibrating process-based models using ground- based hemispherical data and freely available remotely sensed data, highlighting the trade-offs between accuracy and practicality. Third, I evaluate the effectiveness of early-season data-driven yield prediction models across two geographic regions, emphasising the need for region-specific calibrations to maintain accuracy and quantifying accuracy loss due to model transferability. Model performance improved as the season progressed, with Support Vector Regressors achieving an RMSE of 0.23 t ha⁻¹ in the arid regions and Random Forests achieving 0,50 t ha⁻¹ and 0.46 t ha⁻¹ in semi-arid and global models. Fourth, I examine the integration of data-driven models into process-based models through data assimilation techniques, demonstrating how Bayesian assimilation and high-temporal resolution data improve yield prediction accuracy. Bayesian assimilation reduced the prediction errors, decreasing RMSE and MAPE by 25% and 76.5%, respectively, compared to no assimilation approach. This research contributes to the body of knowledge by providing a comprehensive framework for integrating remote sensing data into yield prediction models, supporting precise and timely agricultural decision-making to optimise productivity in water-limited environments.