Summary

The spatio-temporal data from GCM outputs required a certain amount of pre-processing in order to extract suitable data sets that can be applied in the modelling process and there are challenges associated with both the processing and storage for the large scale of such data. Differences in grid resolution for the GCMs posed a challenge in being able to combine multiple models to provide an ensemble average. Selection of suitable grid points requires processing a large volume of raw data in order to result in the final data set. Prior to performing modelling, it is important to establish appropriate evaluation metrics and an approach to numerically determine the uncertainty within the model. Models will be compared across different sites leveraging Kling-Gupta efficiency, Nash-Sutcliffe efficiency, Willmott’s index and RRMSE. The magnitude of error for each site will be indicated by the RMSE and the MAE. A set of baseline models, that have previously been employed in the downscaling task, are developed for comparison, these include the GLM, GLMNET and the Random Forest (RF). Deep learning models consist of a large number of possible configurations, and the modular nature of the architecture requires a search process in order to determine the best configurations based on the training data set. This thesis approached the development of the CNN-LSTM model in multiple stages, where a search procedure was applied with the focus on establishing model parameterisation, defining CNN module architecture, and defining the end to end CNN-LSTM configuration as well as the appropriate learning rate. Once the resulting models are defined, it is possible to evaluate their respective performance on the held out set to determine the relative strengths of each approach.

Chapter 3: Model Development

Summary