Developing gridded climate data using neural networks: high-resolution historical climate and future projections for Africa

• Developing gridded climate data for Africa with deep neural networks

• Author / Creator

  Namiiro, Sarah Aminah

• Databases of high-resolution interpolated climate data are essential for climate change research, such as analyzing impacts of climate extreme events on biological systems and the development of climate change adaptation strategies for managed and natural ecosystems. To enable such efforts, this thesis contributes a comprehensive high-resolution database of historical and future climate data for Africa, including 30-year normal period estimates, decadal averages, annual, seasonal, and monthly data from 1901-2020, as well as CMIP6 multi-model climate projections for the 2020s, 2050s, and 2080s.
  The database contains 48 monthly variables (Tmin, Tmax, Tave, Prec), 16 seasonal variables, and 16 bioclimatic variables (such as degree days, drought indices, etc.) for 142 historical time periods (years, decades, normals) and 168 future projections. A collaborator provided a user-friendly software solution (ClimateAF), with database and the software front-end freely available for download at http://tinyurl.com/ClimateAF, allowing non-technical users to interactively query a total of 24,800 climate grids generated in this study.
  These climate grids are queried with the delta (change factor) method, which is based on a high-quality baseline grid for the 1961-1990 normal period (2.5 arcminute resolution), while all other grids are derived by expressing historical and future data as a lower resolution anomaly layer (0.5 degree), minimizing data storage requirements. The ClimateAF software combines the layers and then further downscales to any desired resolution (up to 250m practically useful resolution in mountainous terrain) with empirical environmental lapse rates.
  The 1961-1990 baseline grid for Africa was developed using a combination of thin-plate spline interpolation of data from 4625 weather stations for Africa, and subsequent fine-tuning with neural networks that associate climate observed at weather stations with covariates reflecting topographic and geographic information (such as elevation, aspect, slope, distance to coast and lakes) in combination atmospheric data from the ERA5 general circulation model (monthly wind direction and strength). The resulting baseline grids accurately model climate phenomena such as precipitation induced by orographic lift on the windward side of mountains, rain shadows, and lake effects on temperature in their vicinity.
  Future climate projections were obtained from 13 Atmosphere-Ocean General Circulation Models (AOGCMs) of the sixth phase of the Coupled Model Intercomparison Project (CMIP6) for four emission scenarios (SSP 1-2.6, 2-4.5, 3-7.0 and 5-8.5) and three future time periods (2020s, 2050s and 2080s). To support users in the selection of a representative set of scenarios for different regions of Africa, I used the Katsavounidis-Kuo-Zhang (KKZ) algorithm which selects an optimally representative set of future projections for 11 regions of Africa, given a user-requested number or scenarios.
  The database was validated using two approaches: To optimize thin-plate spline models and neural network fine tuning, I used a checkerboard validation approach, where the study area was divided into three degree grid cells. Half the cells (“black fields”) were used for model training and the other half (“white fields”) for validation, and vice versa. This approach controls spatial autocorrelations among nearby weather stations and thereby avoids overfitting. The final model was built using all data, and error statistics were assessed using the mean absolute error (MAE) in a non-independent test, including the additional accuracy improvements from ClimateAF’s downscaling algorithms.

• Subjects / Keywords

  • Climate data
  • Interpolation
  • Neural networks

• Graduation date

  Fall 2024

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-jy0v-4728

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.