Integrating crop models and machine learning for projecting climate change impacts on crops in data-limited environments

Context Accurately projecting crop yields under climate change is essential for understanding potential impacts and planning of agricultural adaptation in sub-Saharan Africa (SSA). Crop growth models and machine learning (ML) are often used, but their effectiveness is limited by data availability, precision, and geographic coverage in SSA. Objective This study aimed to integrate ML with a process-based crop model to produce geographically continuous gridded crop yield projections while reducing uncertainties associated with standalone ML or crop growth models. As a case study, we implemented it to project the climate change impact on water-limited potential yield of maize across SSA. Methods We developed an integrated system that combines ML with eco-physiological processes to estimate sowing dates and thermal times, ensuring that crop phenology is accounted for, thus improving potential rainfed yield simulations under varying environmental conditions. Random Forest and crop model-based algorithms are integrated in three steps: (i) RF1, a Random Forest model integrated with a sowing algorithm, designed to estimate the sowing window and sowing date; (ii) RF2, a Random Forest model combined with a crop model algorithm to estimate cumulative thermal time during the growing season, used to determine the timing of phenological stages; and (iii) RF3, another Random Forest model, trained based on eco-physiological principles applied in phases (i) and (ii), employed to simulate water-limited potential yield. The outcomes of the different steps of the framework under historical conditions were tested against reported data across SSA. Results and conclusions For maize and historical climatic conditions, the framework delivers yields which differ less than 20 % of those simulated with a crop model with high-quality inputs, in 95 % of the cases. Our approach thus shows value for generating crop yield projections in data-scarce regions under historical climate, and under future climatic conditions which already feature today somewhere in SSA and for which the framework has been trained. Significance Our approach can also be applied to other major food crops in SSA, under both current and climate change conditions. It allows testing the effect of adaptation of crop cultivars in terms of maturity group. Thus, it can be used for different crops and with far less data requirements compared to process-based crop models. It has the potential for significant applications in assessing climate change impacts, guiding adaptation strategies, and supporting crop breeding programmes and policymaking efforts in SSA.