Machine-learning crop meta-models for climate adaptation

In recent years, computational models have become an important tool in climate impact assessments of agricultural productivity such as crop yields. Presently, such models are typically process-based, which means that various plant and environmental processes are simulated jointly. This level of detail is required to consistently estimate and evaluate various target variables in the plant-environment nexus, for example soil attributes and hydrology besides plant growth itself. However, such models are computationally very costly and challenging to setup and process. To obtain just selected results of their simulations, here crop yields, emulators have become a promising alternative. These mimick the actual model while requiring sparser input data and having a substantially shorter runtime. Besides the time saving per se, this also allows for much more complex scenario analyses. Simple emulators have already been developed in the past but had considerable limitations with respect to spatial resolution and the flexibility of crop management for climate adaptation. This project bases itself on the hypothesis that machine learning methods, i.e., algorithms that can learn from massive amounts of data to provide accurate predictions, facilitate a new generation of emulators that dispel the above limitations and can facilitate complex scenario analyses. In a first step, the project team will produce a comprehensive multi-factorial set of training data using a process-based crop growth model. These data will then be used to train several structurally different algorithms and thereby produce an ensemble of emulators or meta-models. These will be evaluated regarding their individual and combined skills to mimick the original model. Finally, the algorithms will be used to conduct complex scenario analyses, including crop yield predictions for a multitude of climate projections to assess their relevance, the role of spatial resolution for the accuracy of crop yield predictions, and studies of adaptation potentials to climate change through cultivar selection and shifts in growing seasons. These scenarios will provide new insights into the complexity and potentials of dynamic adaptation, which is thus far hardly accounted for in the field of agricultural climate impact assessments. This will be paired with comprehensive assessments of robustness and uncertainty in climate impacts relating to climate projections and spatial resolutions. The algorithms themselves developed within the project are expected to be suitable for a wider range of applications such as near real-time scenario evaluations with agricultural stakeholders

The project Machine-learning Crop Meta-models for Climate Adaptation (MACROS) had the objective to advance the assessment of climate change impacts on global agriculture by combining state-of-the-art process-based crop modelling with machine learning. This allowed to develop fast and highly accurate emulators that can simulate crop yields under a wide range of future climate and management conditions including unprecedented scenario analyses. A major technical achievement of the project was the development of the CROp Model Emulator Suite (CROMES), a new machine-learning framework that emulates complex global crop models with very high accuracy while reducing computational requirements by more than an order of magnitude. This enables researchers to analyze substantially more climate and adaptation scenarios than was previously reasonably feasible. The tool has been published in a scientific journal, and its code and training datasets have been made openly available to the research community. An important application of CROMES was the assessment of uncertainties in climate projections and how they translate into crop yield impacts. Typically, these are based on simulations of up to five climate models and two emission pathways out of more than 30 climate models simulating each four emission pathways for most recent IPCC assessments. By evaluating close to all these climate projections and their impact on crop yields, the project team has shown that standard assessments may overlook relevant future climate conditions, particularly caused by differences in precipitation. Furthermore, the selection of climate models and scenarios plays out differently, depending on whether summer or winter cropping seasons are investigated. The work highlights the importance of considering a broader range of climate futures when planning adaptation and resilience strategies for agriculture. The project has also generated important new findings on agricultural adaptation in double-cropping systems. Research conducted on Brazilian maize-soybean production shows that while individual crops may adapt successfully through altered planting schedules and cultivar choices, the adaptation space becomes more constrained when crops are grown in the same year in a tight sequence. The framework has subsequently been scaled to global applications for monocrops to evaluate where future optimal growing seasons shift most, what the drivers are, and how stable shifts in growing seasons remain. Beyond its scientific outputs, MACROS is already delivering wider impact. The tools developed within the project have been adopted by international researchers for studies on climate impact attribution or broader assessments of climate risks using large ensembles.