Open in app

Sign in

Write

Sign in

Climate adaptation : from model projection to concrete applications with downscaling

Paul Etienne Mallet
Axionable
Published in
15 min readSep 5, 2024
Conceptual view on downscaling. The goal is to generate high-resolution simulations, based on lower-resolution data (source: LSCE)

Introduction

Unprecedented global temperatures led 2023 to become the warmest year on record, while 2024 will probably surpass it, as the warmest day record has been beaten several times this year. In the near term, every region in the world is projected to face climate change impacts such as modifications on human activities and climate hazards. In this context, assessing future evolutions of climate and its consequences becomes increasingly important for adaptation.

Climate model projections are the major tool to do so. However, decision-makers investigating how infrastructures, agricultural yield or energy will suffer from weather modifications in the next decades will not be able to use them directly. Indeed, their spatial resolutions are too coarse: one of their cells covers areas typically as large as a country like Lebanon, or even the Netherlands. A location situated in a rainy oceanic area would be considered with the same weather conditions as one hundreds of kms farther in a dry mountain. To solve this issue and considerably enhance climate projections usefulness, climatologists have developed spatial disaggregation techniques, also known as downscaling.

Downscaling consists of a set of techniques aiming at estimating variables at a finer scale than the available ones. People interested in climate and related impacts, risks, and opportunities will probably encounter downscaled data and would benefit from familiarizing themself with main concepts and methods. Moreover, climate data users may want to produce their own downscaling to fit their specific use cases, as a mismatch between the needed data resolution and the available one is indeed very common.

In this article we give an overview of the large ensemble of techniques aimed at solving this problem with a focus made on future climate projection spatial disaggregation. In addition, other downscaling applications are discussed, such as remote sensing data and temporal disaggregation. Finally, a post-processing methodology employed to improve climatic data quality, known as bias correction, is presented.

Comparison of model resolution impacts to describe reliable weather. Average of precipitation in winter over UK view by models from 300 to 25 km resolution compared with MetOffice observations at 8 km (source: Maraun et al. Precipitation downscaling, 2010).

Main use case: climate projections

The coarse resolution of global climate models (GCMs) to estimate climate change in the future is one of the main application cases for downscaling techniques. Climate models are numerical simulations of the atmosphere. They consist of a virtual subdivision of the atmosphere into a 3D grid, each cell of the grid representing an air parcel by its average characteristics. At every time step, the evolution of these characteristics is calculated with the physical equations valid for the atmosphere. To make reliable projections of future climate, it is necessary to start from the most complete representation possible of the climate system. This means having a global model, i.e. covering all regions over the world, but also all climatic components and not only the atmosphere. Thus, the models used for climate projections are atmospheric models coupled with other complementary models representing the ocean, ice or land surfaces, and integrate increasingly complex representations of atmospheric physical and chemical processes.

Conceptual structure of a climate model. (source: NOAA, 2012)

Such a complex simulation involves a huge number of parameters to be calculated at each time step, and this over decades. Consequently, to limit the IT costs, climate projections are made on very coarse resolutions, with a typical mesh size of 100 to 200 km. Concerning the consequences of climate change over a given territory, the information provided by the GCMs will be very likely to be unusable directly. Indeed, atmospheric variables will be aggregated over vast regions that do not distinguish all sub-grid heterogeneities, such as reliefs, land and sea surfaces and other characteristics of the territory. Moreover, the sub-grid physical and chemical processes will not be directly simulated from the atmospheric physics equations, but parameterized in the GCMs, leading to errors in the representation of local phenomena, especially in the estimation of extremes.

Dynamical versus statistical downscaling

Two very different main approaches are used to meet the need for fine-resolution data from low-resolution data.

Dynamical downscaling

The first approach consists in using a second model to estimate atmospheric fields at a finer resolution from their coarse description obtained as output of the GCMs. This second model is a regional climate model (RCM), its resolution is typically of 10 to 50 km and its utilization is limited to a selected area instead of the whole world. In concrete terms, the RCM is responsible for simulating at high resolution the atmosphere as it is described at low resolution by the GCM in a dynamical way by computing evolution of the situation at each time step. To do so, the RCM is forced to meet the atmospheric state as it is described at coarse resolution by GCM. In the most widespread technique, called lateral boundary forcing, the RCM is free to compute evolution of the atmosphere without constraints inside its area if the atmosphere characteristics meet their GCM large-scale representation at the edge of the domain. By consequence, in the context of climate projections, the RCM does not directly simulate climate modifications under climate change, but rather simulates what the GCM’s coarse variables would correspond to at a finer scale, estimating the future climate.

Schematic representation of dynamical downscaling (source: Deutscher Wetterdienst)

Pros of dynamical downscaling:

  • Some physical processes that GCMs are not able to simulate because they are sub-gridded, are calculated here. Therefore, downscaling has a good chance of improving physical representation of phenomena.
  • The RCM chosen to simulate a given region can be fine-tuned to a better representation of the processes of the region concerned.

Cons of dynamical downscaling:

  • Some of the results are biased, as models are limited in their ability to realistically simulate all atmospheric processes.
  • Very expensive in terms of IT resources
  • And above all, you need to have modeling capabilities, which involves a long development with a high degree of expertise in atmospheric physics, but also supercomputing possibilities, which is generally reserved for specialized research laboratories and institutions. In addition, calibrating and validating a RCM to simulate the climate for a given area is a tedious work, which generally lasts one or two decades for experimented teams.

While it is very useful to understand dynamical downscaling when working with climate data, this last condition means that simple users have no chance of applying this technique themselves to meet their needs.

Dynamical downscaling: The CORDEX example

A representative example of downscaled data comes from the CORDEX experiment (COordinated Regional climate Downscaling EXperiment), an international coordination effort for regional projections. To make climate projections more usable for concrete use cases, the CORDEX project proposes downscaling of GCM simulations for a set of 14 large regions covering almost all land areas of the world. Critical analysis by experts makes this dataset even more relevant, for example with a selection of the best-performing GCMs for each region, or a comparison of the performance between the different RCMs for each region.

Boundaries of the fifteen CORDEX domains on which GCM climate projections are downscaled (source: CORDEX-CMIP5)

Statistical downscaling

A conceptually very different approach for spatial disaggregation is the statistical/empirical one. Here, high-resolution data is estimated not by simulating the underlying physical processes as in dynamical downscaling, but by using the statistical relationships observed between the large and small scale.

In concrete terms, two datasets covering the same historical period are considered, one at a large scale and the other at a local scale, with the aim of establishing a correspondence between local variables and the large-scale atmosphere. To represent the local variable, data from observations or outputs from climate reanalysis are generally used. Once this relationship has been established over a period long enough to be statistically robust and to have covered a maximum of atmospheric configurations, it is used to disaggregate large-scale data to a high resolution in a context where the local value is not available.

Statistical downscaling adjustment illustration (source Jet Propulsion Laboratory)

Pros of statistical downscaling:

  • A huge advantage of this solution is that it is much less expensive in terms of resources than Dynamical downscaling.
  • Accessible to users with little experience. It can be implemented at different degrees of complexity depending on the time available, the need or the level of expertise, from very simple linear transformations to advanced techniques.
  • Flexible. Technical solutions can be adapted accordingly to the issue.
  • Adapted to specific location: unlike dynamical downscaling, which results in data projected onto a grid, this time the information can be downscaled at specific coordinates, such as observation stations for example.
  • It takes into account implicitly local conditions, such as topography or vegetation. Whereas dynamical downscaling has to make complex assumptions on parametrization to consider them, those are implicitly included in the historical behavior of the observed variables.
  • Finally, it avoids some of the biases induced by dynamical downscaling. Indeed, the statistical approach is in many situations more reliable than the physical process resolution.

Cons of statistical downscaling:

  • One of the main disadvantages of this solution is that it is based on fixed relationship between large and small scales, that can actually evolve over time, especially due to climate change.
  • Time coverage dependent. If the reference period is too small to establish the relationship, some extremes may not be represented.
  • Certain physical behaviors are missed by the statistical approach, while the physical resolution by RCM is able to represent them.
  • May create unrealistic situations. The first type of possible inconsistency is spatial: the statistical downscaling carried out towards a series of localities does not consider the relationships between these localities. Thus, a variable whose values at 2 nearby stations are strongly correlated with each other in reality, may appear to be uncorrelated with downscaling. The second type of inconsistency is physical. The behavior of the different variables representing a parcel of air are not independent of each other but related by physics. Roughly speaking, the law of perfect gases states that the pressure of a parcel of gas varies in proportion to its temperature, and inversely proportional to its volume. Thus, the statistical downscaling of these three variables independently is likely to create situations in which they vary independently of each other, which has no chance of real existence.

While dynamic downscaling requires long developments and is reserved for specialized institutions, statistical downscaling is much more accessible. Its optimal utilization may require advanced knowledge, but it is possible to have very satisfactory results by implementing affordable solutions.

The main families of statistical downscaling techniques

Many statistical downscaling techniques exist, which consist most of the time of establishing an empirical transfer function to link large-scale atmospheric variables as predictors to local-scale ones as predictands.

To classify them, a first criteria is the type of predictor used:

  • Perfect Prognosis (PP) uses observations as predicators to calibrate the relationship
  • Model Output Statistics (MOS) uses model outputs as predicators to calibrate the relationship.

This simple difference generates some important implications. In particular, PP models are supposed to be physical based, in the sense that the divergence of the variable properties when passing from the large to the local scale is considered to rely only on physical scale effects, such as topography implications and other local meteorological conditions, while modeling assumptions are not considered. By consequence a PP model, once calibrated, is supposed to be applicable to any GCM. In contrast, MOS produces GCM-dependent transfert functions, which can be seen as a disadvantage since a new transfert function should be trained for each GCM, but also an advantage since the very specific GCM-dependence of MOS models produces more accurate downscaling, which means that this method also involves model bias correction.

A second criteria to classify those techniques is the type of transfer function considered. This aspect meets a more common aspect when dealing with data: the type of statistical approach. Generally, they are divided into two categories:

  • Parametric functions
  • Non-parametric functions

A parametric function is a mathematical function that is defined by a set of parameters. When using this type of function, we assume the fact of having a pre-established conception of the type of relationship existing between the large and local scales. The simplest type of parametric function is obviously the linear function. In the context of downscaling, using it involves assuming that the local-scale variable behavior is a linear representation of its large scale, which can be an acceptable approximation and is widely used. A more evolved and adaptable function family is Generalized Linear Model (GLM) and its extension, the Generalized Additive Model (GAM). Parametric functions present advantages of being computationally efficient and interpretable.

On the other hand, a non-parametric function does not rely on a pre-specified mathematical equation with a limited number of parameters but is instead data-driven. It can capture complex relationships in data that parametric models might miss and offer good performance in case of limited prior knowledge about the data. A generalized approach is quantile-mapping, consisting of matching quantiles of the large-scale distribution to corresponding quantiles of the local-scale distribution. This transformation corrects not only the mean and variance but the entire distribution, including extreme values. During recent decades, machine learning techniques, such as Support Vector Machines (SVM) or Artificial Neural Networks (ANNs), have been tested to provide transfer functions and are more and more adopted.

In addition to that, two other approaches can be included:

  • Analog method: each local-scale variable value is relied to the most similar large-scale weather situation in the past based on a metric such as Euclidian distance. In other words, with this approach the local values are linked not only to the variable value at large-scale, but to a much larger condition: the synoptic atmosphere circulation state.
  • Weather Generator (WG): a random generator of synthetic weather data for a given site, which has statistical properties similar to the time series observed at that location. Here it will not be a question of identifying the most likely value of the local variable corresponding to a particular large-scale condition, but instead of stochastically creating a variable that respects the characteristics of statistical distributions observed locally. WGs are mainly used for downscaling precipitation. Indeed, due to the peculiar distribution of precipitations, far away from gaussian but instead binary-like shaped, the local behavior of this variable is more difficult to reproduce by common technics. Furthermore, they have the capacity of providing unlimited length time series.

Existing statistical downscaling tools and libraries

To facilitate downscaling handling, scientists and specialized institutions have shared some tools. Most are in R, while a few are in Python:

Other use cases

Climate data downscaling is presented here for its main application, the spatial disaggregation of climate model outputs. However similar approaches for other use cases should be mentioned and are treated separately because of the specificities of their purpose and relative techniques. These concern post-processing of remote-sensing observations and temporal disaggregation. Furthermore, bias correction is also discussed because of the close ties with downscaling.

Remote sensing data fusion

Earth and climate observations through remote sensing methods can provide information on phenomena from a wide range of points of view, such as different spatial and temporal resolution or different spectral coverage. A remote sensing specific downscaling method is then to estimate a variable at high resolution through the fusion of, for example, a high-resolution observation data with a weak temporal coverage and a low spatial resolution data with a good temporal coverage. The most advanced technique is today to use deep learning models, mainly convolution neural network (CNN) or generative adversarial network (GAN).

Illustration of spatiotemporal data fusion (source: Spatiotemporal Fusion of Multisource Remote Sensing Data: Literature Survey, Taxonomy, Principles, Applications, and Future Directions, Zhu et al., Remote Sensing, 2018)

Remote sensing data fusion techniques are also used to combine the different properties of the physical object captured by different sensors. As an example, radar imaging can add optical imaging information on surface topography and can cover areas hidden by clouds or volcanic smoke.

Temporal disaggregation

Until now, we have talked about spatial downscaling, to cover the need for data with a higher resolution than the available data. Temporal disaggregation consists of doing the same for the time dimension. For example, a user needing data sampled on an hourly time step can consider using daily sampled data with some adjustment if this is the only available.

The first approach is to reconstruct natural cycles around aggregated values. For example, as a first guess, it’s possible to pass from a daily to an hourly sampled temperature time series by adding a sinusoidal modulation component counting for diurnal cycle.

The results obtained with this method are rather good for variables presenting smooth curves and easily recognizable cyclic behavior. However, for other variables, and in particular for precipitation, this technique is useless. Indeed, precipitation time series patterns are much more binary (rain or no-rain) and look like a succession of peaks with various intensity and duration. Because of the large impact this sub-daily distribution has on precipitation issues, this aspect is widely discussed in the community and lot of methods have been developed, and stochastic weather generators are among the most popular.

Downscaling and bias-correction

A bias in climatic dataset is a systematic error on the estimated environmental variable. Both climate models and remote sensing instruments present bias in their outputs. Those errors can be due to a wide range of factors. Concerning climate models, the simplifications made on physics for reducing computing cost, or simply an incomplete understanding of every climate system aspects are involved.

When dealing with climate data it is highly recommended to apply some bias adjustment to enhance the confidence of data quality. To do so a post-processing step is applied before utilization. The methodologies developed for bias correction or statistical downscaling are in many cases very close to each other. They consist both in the calibration of a statistical transfer function based on observations as reference dataset, and have common issues, as the representation of extremes.

More precisely, this step can be done after the application of the downscaling process or can be included in this process itself. Indeed, in the MOS approach the downscaling operation includes a bias correction process, as the calibration of the transfer function aiming to fit observations is done for a specific model, taking into account its systematic errors.

Conclusions

This article presents an overview of downscaling methodologies, and in particular its most common use: the regionalization of future climate projections. This one is mandatory in majority of cases to reliably assess how climate change impacts human activities, which is becoming critical as an increasing number of businesses and populated areas are already concerned and further dramatic consequences are expected.

The spatial disaggregation of GCM projections is performed either by dynamical or statistical downscaling. The dynamical one requires significant resources and is reserved for specialized institutions. Statistical downscaling is much more affordable, and its simplest techniques can be performed quite easily. Its quality will be highly dependent on the available data. These must be of high quality, cover long periods of time including as many weather configurations and extremes as possible. The confidence that can be placed in the results should be assessed by calculating uncertainty on historical data.

The downscaling approach have been outlined in this article and more specific aspects will be treated in upcoming ones, concerning a more advanced statistical downscaling method description, bias-correction techniques as those utilized to produce CORDEX data regionalized over France presented by the DRIAS platform, or a coding example of temporal disaggregation of daily to hourly sampled data.

Bibliography

Downscaling GCM:

  • A Review of Downscaling Methods for Climate Change Projections, S. Trzaska and E. Schnarr, 2014
  • The impacts of climate change on extreme rainfall and flooding of Mediterranean mesoscale watersheds, A. Colmet-Daage, Thesis (in french), 2018.
  • Precipitation downscaling under climate change: recent developments to bridge the gap between dynamical models and the end user, Maraun et al. 2010
  • pyESDv1.0.1: An open-source Python framework for empirical statistical downscaling of climate information, Boateng and Sebastian G. Mutz 2023

CORDEX:

  • Projected future changes in tropical cyclones using the CMIP6 HighResMIP multimodel ensemble. Roberts, M. J., Camp, J., Seddon, J., Vidale, P. L., Hodges, K., Vannière, B., et al. Geophysical Research Letters, 2020
  • Regional Dynamical Downscaling and the CORDEX Initiative, F. Giorgi and W. J. Gutowski Jr., Annual Review of Environment and Resources, 2015

Remote sensing downscaling:

  • A review of spatial downscaling of satellite remotly sensed soil moisture, Peng et al. 2017
  • Spatiotemporal Fusion of Multisource Remote Sensing Data: Literature Survey, Taxonomy, Principles, Applications, and Future Directions, Zhu et al., Remote Sensing, 2018

Bias correction:

  • Ensemble bias correction of climate simulations: preserving internal variability, Ayar et al, 2021
Downscaling
Climate Change Adaptation
Resolutions
Climate Data Science
Bias Correction

--

--

Paul Etienne Mallet
Axionable

Written by Paul Etienne Mallet

0 Followers
Writer for

Help

Status

About

Careers

Press

Blog

Privacy

Terms

Text to speech

Teams