Convolutional Neural Networks for Climate Downscaling

A key challenge in climate modeling is the assessment of the impact of global climate variables on regional weather measurements such as temperature and precipitation. This assessment is usually done by downscaling the output of (coarse resolution) global climate models to regional (high resolution) predictions. There are two independent downscaling pathways: dynamic and statistical. Dynamic downscaling uses high resolution physical models of regional climate in conjunction with global models to make regional climate predictions. Although dynamic downscaling methods account for regional geographic variations, they are computationally expensive even on modern supercomputers. Statistical downscaling algorithms, on the other hand, establish correlations between global model outputs and historical (decadal) weather observations, developing statistical models that translate global model output into regional weather projections. In recent times, many machine learning (ML) algorithms and statistical methods have been used for statistical downscaling, e.g., Bayesian frameworks, artificial neural networks, support vector machines, multilayer neural networks [4, 5, 7, 9] and ensemble methods (e.g., boosting, bagging and stacking).However, climate downscaling poses some open challenges such as: incomplete representation (and non-linearity) of physical processes in global models, non-stationarity in the global models and (hence) the relationships with regional observations, and computational complexity of processing large Petabyte-scale datasets. Many existing ML algorithms are limited in their ability to model complex functions with many variations that represent deep relationships between input and output variables [1]. Research in ML shows that architectures designed to model complex functions need to be of sufficient depth (i.e., maximum length of path from any input to any output) to prevent poor generalization [2]. Many ML algorithms do not correspond to deep architectures, e.g., under certain assumptions, decision trees have two levels, logistic regression has one, kernel machines have two levels and ensemble methods add one level to the base learner [1]. Even in multilayered neural networks, learning algorithms perform poorly when complex functions require (error) gradients to be propagated across many levels. Another limitation of many ML algorithms is the use of local estimators, i.e., input space is partitioned into regions and the degrees of freedom in the underlying model are used to account for variations of the target function in these regions. As a result, algorithms such as decision-trees and local kernel machines (e.g., Gaussian processes and manifold learning algorithms) may not generalize well for complex functions with many variations [3].