Search for a project

Using deep machine learning and crop models to inform adaptation

Prof Andy Challinor (SEE), Dr Edward Pope (Met Office), Prof Anthony Cohn (School of Computing), Prof Netta Cohen (School of Computing)

Project partner(s): Met Office (CASE)

Contact email: A.J.Challinor@leeds.ac.uk

Overview

Climate change presents numerous challenges to society. When the fifth assessment of the Intergovernmental Panel on Climate Change (IPCC) was released in March 2014, food production and food security was highlighted as a key area of concern (see e.g. Guardian article). Adaptation options for crop production systems are assessed using crop simulation models together with climate models.

Process-based crop models have been used to develop options to adapt to climate change (e.g. Webber et al., 2014, Challinor, 2009). Statistical models of yield responses to weather and climate have also been used to develop adaptation options (Lobel and Burke, 2010); and it is not yet clear which models are best, even for specific appliations (Estes et al., 2013). Crop-climate indices are a broader set of tools that focus on specific biophysical processes, rather than attempting to reproduce yields. Again, evaluation is difficult; however the technique is clearly powerful, especially when used in conjuction with process-based modelling to answer specific research questions on adaptation (Challinor et al., 2016).

Machine Learning techniques (ML; see Witten et al., 2017) present a more powerful set of tools than the narrower and simpler statistical models and crop-climate indices. This PhD use ML techniques as an integrative tool for multiple methods of assessing climate impacts on crops. The ultimate aim is to use those methods to identify adaptation options.

Novel Methodology

The PhD will explore inform next generation of crop models by combining ML with process-based models. The underlying research questions for this aim are:

  1. Skilful prediction: how much data is needed for deep machine learning to outperform current process-based crop models?  
  2. Improved understanding: can explicit spatial knowledge about fundamental biotic and abiotic limitations to crop productivity be extracted from deep learners? Can current spatial variation in climate be used as analogues for future climates?
  3. Improvement of methods: can ML analysis of GCM data lead to improved indices of crop response to current and projected climates? At which spatial and temporal scales should indices be developed in order to remain robust when used with projected (especially out-of-sample) climate data?
  4. Integration of methods: how can machine learning and modular use of crop models be used to identify common modes of behaviour across process-based crop models? Can one model be used to emulate a range of other models?

Techniques

Crop modelling has been used for various applications over the past few decades. Field-scale applications for decision support have a long history (Hoogenboom et al., 1994) that in turn enabled frameworks to link crop and climate models (Challinor et al., 2003). The need to quantify uncertainty (Challinor et al., 2013) has led to work supporting the use of crop models with climate ensembles (Ramirez-Villegas et al., 2013).

 

Climate models are used around the globe to assess climate impacts. Much of the current food security work at the Met Office is focussed on understanding the influence of natural climate variability and change on crop yield variations. Indicators have been developed using observations over relatively large spatial scales (e.g. province/state level) for use with climate models.

Machine learning is now a ubiquitous technique and methodology in many fields, including big data, health, environment, robotics etc. There is a huge demand by both academia and industry for PhDs with machine learning expertise. Of particular relevance has been the recent success of “deep learning” (as demonstrated by the success of Google’s AlphaGo in beating a world expert at the game of Go, a challenge still thought to be many years from achievement). Deep learning systems are based on the earlier neural network learning systems, but differ in that deep learning systems have many more “hidden layers”, which allow the system to learn intermediate representations and features which allows for a more expressive representation and better performance. One requirement for success is that a sufficient amount of training data be available; if this proves problematic there are many other machine learning methods which could be investigated.

The advantage of applying machine learning to crop modelling is that potentially allows the use of much more accurate models, and which do not require intensive human intervention to produce; potentially they can also adapt to changing conditions. Depending on the machine learning method finally adopted, it also allows introspection by human experts of the model learned.

Combining crop-climate modelling with ML is relatively new. Progress has tended to focus either on index-based modelling of crop and weather responses at regional scales (Chavez et al., 2015) or genomic / small scale approaches (e.g. Heslot et al., 2014). Use of ML with process-based crop-climate modelling, alongside integrated assessments, is expected to be an important part of the next generation of crop impacts and adaptation assessments.

Training and research environment

The student will be based in the School of Earth and Environment (SEE) at Leeds, with more than 220 fellow PhD students. He/she will benefit from the School’s first class research facilities and from outstanding job prospects in government, academia, research and industry. The student will be a member of the Climate Impacts Group, an active team of researchers working on developing and using climate and crop models to quantify the impacts of climate variability and change on crop yield, including associated uncertainties and adaptation options. The group meets twice a week: once for scientific exchange, and once for an informal lunch.

The student will be co-supervised by Professors Cohen and Cohn from the School of Computing who bring computing and specifically machine learning expertise to the supervisory team. The student will also have access to facilities in the School of Computing and to interact with the students there, and to participate in mentoring and training events in Computing.

The industrial partner for this studentship is the Met Office. The student will be an important part of the wider Met Office team that brings complementary expertise and skills to core climate and impact science done under the Met Office Hadley Centre Climate Programme (HCCP). Food security is a key priority for the Met Office Hadley Centre, building on and using state-of-the-art climate observations, projections and understanding of the climate system in developing new crop indices and improved modelling techniques, which feedback to the core research programme. The student will attend all the relevant meetings that are held to direct the science and will have close working ties with Met Office staff and relevant stakeholders, thus developing links to potential future employers. The student will spend at least 3 months at the Met Office, as several visits of 3 days or more, spread across the project duration to cover general training and induction, model specific training, and discussion of results and paper preparation.

The studentship will form an integral part of both the Leeds-York DTP and the Met Office Academic Partnership. Both of these will provide significant cohort interaction.

Training will be provided to cover familiarisation with the complex modelling tools and data involved. PhD students at Leeds also undertake a 2-day induction course, which includes an overview of regulations relating to research degree students; health and safety briefing; skills training, and departmental finance guidelines.  Students complete with their supervisor a training needs assessment within 1 month of starting. They manage their research training using an on-line Personal Development Record where all training courses and monthly meetings are uploaded. This is used to plan and reflect on what skills are being developed.

References

CHALLINOR, A. J., KOEHLER, A. K., RAMIREZ-VILLEGAS, J., WHITFIELD, S. & DAS, B. 2016. Current warming will reduce yields unless maize breeding and seed systems adapt immediately. Nature Clim. Change, 6:954-958, doi:10.1038/nclimate3061

CHALLINOR, A. 2009. Towards the development of adaptation options using climate and crop yield forecasting at seasonal to multi-decadal timescales. Environmental Science & Policy, 12, 453-465. https://doi.org/10.1016/j.envsci.2008.09.008

CHALLINOR, A. J., SMITH, M. S. & THORNTON, P. 2013. Use of agro-climate ensembles for quantifying uncertainty and informing adaptation. Agricultural and Forest Meteorology, 170, 2-7. https://doi.org/10.1016/j.agrformet.2012.09.007

CHALLINOR, A. J., WATSON, J., LOBELL, D. B., HOWDEN, S. M., SMITH, D. R. & CHHETRI, N. 2014b. A meta-analysis of crop yield under climate change and adaptation. Nature Climate Change, 1-5. doi:10.1038/nclimate2153

CHAVEZ, E., CONWAY, G., GHIL, M. & SADLER, M. 2015. An end-to-end assessment of extreme weather impacts on food security. Nature Clim. Change, 5, 997-1001. doi:10.1038/nclimate2747

ESTES et al. (2013).Global Ecology and Biogeography, (Global Ecol. Biogeogr.) 22, 1007–1018. DOI: 10.1111/geb.12034

HESLOT, N., AKDEMIR, D., SORRELLS, M. E. & JANNINK, J.-L. 2014. Integrating environmental covariates and crop modeling into the genomic selection framework to predict genotype by environment interactions. Theoretical and Applied Genetics, 127, 463-480. doi:10.1007/s00122-013-2231-5

HOOGENBOOM, G., JONES, J. W., WILKENS, P. W., BATCHELOR, W. D., BOWEN, W. T., HUNT, L. A., PICKERING, N. B., SINGH, U., GODWIN, D. C., BAER, B., BOOTE, K. J., RITCHIE, J. T. & WHITE, J. W. 1994. Crop Models, Honolulu, Hawaii, University of Hawaii, Department of Agronomy and Soil Science.

LOBELL, D. B., & BURKE, M. B. (2010). On the use of statistical models to predict crop yield responses to climate change. Agricultural and Forest Meteorology, 150(11), 1443–1452. https://doi.org/10.1016/j.agrformet.2010.07.008

RAMIREZ-VILLEGAS, J., CHALLINOR, A. J., THORNTON, P. K. & JARVIS, A. 2013. Implications of regional improvement in global climate models for agricultural impact research. Environmental Research Letters, 8, 024018. doi: 10.1088/1748-9326/8/2/024018

WEBBER, H., GAISER, T. & EWERT, F. 2014. What role can crop models play in supporting climate change adaptation decisions to enhance food security in Sub-Saharan Africa? Agricultural Systems, 127, 161-177. http://dx.doi.org/10.1016/j.agsy.2013.12.006

WITTEN, H., Frank, E. Hall, M. A., and C. J. Pal. Data Mining, Fourth Edition: Practical Machine Learning Tools and Techniques (Morgan Kaufmann Series in Data Management Systems). See http://www.cs.waikato.ac.nz/ml/weka/book.html

Related undergraduate subjects:

  • Atmospheric science
  • Computer science
  • Earth system science
  • Engineering
  • Environmental science
  • Mathematics
  • Meteorology
  • Physics
  • Plant science