Introduction

Downscaling of climate change models is the procedure of using large-scale climate models to
make climate predictions at finer temporal and spatial scales to fit the purpose of local level
analysis and planning (UN-CTCN, 2024). The large-scale climate change models can be used to
study impacts on natural resources, human health, commerce and infrastructure. Downscaling
involves transforming coarser resolution of global climate models, GCM, to finer resolutions
spatial scale data. It aims to transform the GCM model data closer to the station level data, in this
case, closer to the small-holder farmer’s needs. Therefore large scale modeled or observed data
can be narrowed down to local or regional scale information. Downscaling procedures include
the addition of certain geographic details into a model, for example, water bodies, presence of
mountains, forest and urban areas. This creates a model that can emulate weather-related
processes that affect local environs like small-holder farms.

Downscaling typically involves use of Global Climate Models, GCMs that represents physical
processes in the atmosphere, ocean, land surface, and cryosphere, which simulate the response of
global climate systems to increasing greenhouse gas emissions and concentrations, using
different emissions scenarios.

Some tools used in downscaling climate include the following:

1. Regional Climate Model (RCM). This uses a high resolution model to derive smaller-
scale information, for example, for small-holder farms. This is the dynamical down-
scaler.
2. MicroMet: This is a spatial and temporal downscaling model that produces high-
resolution climate data. MicroMet uses ground-based observations to calculate high-
resolution gridded climate data
3. KrigR: This tool allows for downloading and statistical downscaling of climate data.
4. SDSM 4.2: This tool allows for decisions by users by comparing two data sets, such as
observed data and downscaled data.
5. PLS-Regression: This tool is used in conjunction with statistical downscaling, example,
the KrigR tool, to reduce bias

1
 6. Global Climate Models (GCMs): These are the primary tools used in generating weather
and climate. They are 3-dimensional computer models that simulate human-induced
climate change.

Invariably, downscaling of climate data is the derivation of local to regional-scale (10-100km)

information from larger scale modeled or observed data (500-10000km). There are two main
approaches to achieving this, which are the dynamical downscaling and the statistical
downscaling. And the main tool used is the General Circulation models, GCMs, which are
computer representations that mathematically show various physical processes of global climate
system.

However, GCM results should only be considered at global or continental scales for climate
conditions averaged at monthly, seasonal, annaual and longer time scales because of the
limitations on computing resources and input data, even through these are well known processes.

Any information that is presented at spatial scales finer than 100km x 100km and temporal scales
finer than monthly values has undergone a process called downscaling, and involves additional
information, data, and assumptions, which leads to further uncertainties, and limitations of
results. Downscaling is therefore the designing and interpreting of fine-scale information on
climate change and its impacts.

The diagram below shows the components involved in developing global and regional climate
projections.
Socioeconomic
storyline

Emissions
Climate Forcing
scenario

GCM
Climate simulation

Climate
projection

Observed RCM
climate 2

Statistical Dynamical
downscaling downscaling
Downscaling relies on ;

Source:
1. TheDaniels et al., 2012
assumption that local climate is a combination of large-scale atmospheric features
and local conditions like topography, water bodies and land surface properties.
2. Deriving climate projections at local scales is a multistep process (Figure above).
Uncertainties creep in and are inherent, arising from different sources and should be
adjusted.
3. Downscaling can be applied to both spatial and temporal data, and many methods are
used to obtained the desired climate change and information.
4. Two principal methods are used: Dynamical, which explicitly includes additional data
and physical processes in models like the GCMs both at a higher resolution, covering
select portions of the globe; and statistical, which shows the relationship between large-
scale climate features that GCMs and local climate characteristic provide. These are easy
to implement and interpret compared to the dynamical method, where large computing
power is required. However the statistical required historical records.

In order to understand downscaling, whether dynamically or statistical, one must understand the
GCMs. The GCMs is the source of downscaling. The spatial resolution of the GCMs is quite
coarse, and has a grid cell size of about 100-500km, with each cell being homogeneous, and the
temporal scale is of monthly means and longer. Therefore, the GCMs provide values that
estimate future climate change at global and continental scale and over long periods. On the
other hand downscaled climate data does the opposite, i.e, provide a finer-scale resolution of
smaller sized cells, usually 10-50km, with the cells being heterogeneous, and temporal scale of
short length, say a week. This means that the GCMs are the source for downscaling, whether
spatially or temporally, for example, coarser-resolutions GCM of 500km grid cell can be
downscaled to a 20km fine-scale resolution, or even specific locations, like our small-holder
farmer’s farm or locality. Dynamical downscaling requires the use of a Regional Climate Model,
RCM, similar to the GCM in its usage but of a higher resolution. The RCMs use the large-scale

3
GCMs atmospheric information at the lateral boundaries and adds more complex topography,
land-sea contrast, surface difference and detailed features of physical processes in order to
generate realistic climate information at a spatial resolution of about 20-50km, thereby reducing
the GCMs resolution of 100-500km. Therefore, the quality of RCMs is tied to forcing of GCMs,
and are subjected to systematic errors that need corrections as well as further downscaling to a
higher resolution.

“Many of the processes that control local climate, e.g topography, vegetation and hydrology, are
not included in coarse-resolution GCMs. The development of statistical relationship between the
local and the large-scales may include some of these processes implicitly” . Viner, 2012.

4
Source: Daniels et al., 2012

Statistical downscaling requires the merging of empirical relationships between historical and/or
current large-scale atmospheric and local climate variables. Once the relationship is established
and validated, future atmospheric variables that GCMs project are used to predict future local
climate variables. Statistical downscaling can provide site-specific climate projections, which
RCMs cannot due to their computational limitations of 20-50km spatial relationships. However,
this approach relies heavily on the assumption that relationship between present large-scale
circulation of the atmosphere and the local climate conditions remain valid under different
forcing conditions of possible future climates (Zorita et al., 1999).

Often times both approaches are used in conjunction. A dynamical – statistical downscaling
involves the use of an RCM to downscale GCM output before statistical equations are used to
further downscale RCM output to a finer resolution. Dynamical approach improves specific
aspects of regional climate modeling and provides better predictors for further statistical
downscaling to higher-resolution output.(Guyennon et al., 2013).

But statistical-dynamical downscaling is a more complex approach that is also less

computationally demanding in comparison to dynamical downscaling. This method statistically
pre-filters GCM outputs into few characteristic states that are further used in RCM simulations.
(Fuentes and Heimann, 2000).

When using dynamical, statistical or a combination of the two, uncertainties come into play.The
GCMs project opposite changes in rainfall amount in certain regions, which shows uncertainly of
future climate change projections even with state-of- the- art GCM tools used. Uncertainty
comes from four main sources.

1. Uncertainty in future levels of anthropogenic emissions and natural forcing , for

example, volcanic eruptions.
2. Uncertainty linked to imperfect model representation of climate processes.
3. Imperfect knowledge of current climate conditions that serve as starting point for
projections; and

5
 4. Difficulty in representing internal annual and decadal variability in long-term projections
(Daniels et al., 2012).

However, uncertainty does not mean that future projections cannot be made. Uncertainty does
not mean that future projections are unknown or false. Uncertainty can be quantified, and
therefore decisions can be made even in the face of the uncertainty.

Downscaling Approaches:An Overview Dynamical Downscaling: Refers to use of an RCM

driven by a GCM to simulate regional climate. RCM differs from GCM in that it has a higher –
resolution and additional regional information that enables it to better represent local landscape
and possible local atmospheric processes. (CTCN, 2024).

The coarse resolution of the GCM need to be approximated to get a clearer finer representation
of the global circulation via a larger numbers of processes. This is where the RCM comes into
play, where at the resolution of 25 – 50km (unlike GCMs 100-500km) for parts of the globe, the
RCM is able to capture some of those smaller-scale processes more realistically. It is able to do
this by feeding the atmospheric fields like surface pressure, temperature, wind and humidity
simulated by GCM into the RCMs vertical and horizontal boundaries. See diagram above.

Locally specific data and physics–based equations are then used to process this information and
obtain regional climate outputs. RCMs can model atmospheric processes and land cover changes
quite explicitly.

Assumptions of RCMs: Dynamical Downscaling

(1) The challenges of RCMs is the large number of computational requirements, even larger
than GCMs.
(2) They also need a lot of inputs and also need complex calibration procedures to make
realistic simulations of atmospheric conditions.
(3) RCMs also have difficulty in simulating convective precipitation accurately.
(4) They cannot also accurately simulate extreme precipitation, and can worsen when
resolution is increased. This is the systematic bias that has to be corrected to get better
information. (Brown et al., 2008).

6
 (5) It also depends on the driving GCM information, for example, if the GCM misplaces
storm tracks, errors show in the RCM. RCM grid boxes are usually greater than 10km,
making it too coarse for hydrological and agricultural impact studies.
(6) For higher resolution results, statistical methods are used in lieu of RCM, or RCM output
is further downscaled via statistical means. (Benestad, 2009).

RCMs and Application:

RCMs are developed by research institutions having enough computational capacity and
technical expertise. Various RCMs differ in numerical, physical, and technical aspects. The
common types are: U.S RCM (RegCM3): Canadian RCM: UK Met Office (HadRM3): German
(REMO): Dutch (RACMO): and German HIRHAM, which combines the dynamics of the
HIRLAM and ECHAM. All these types of RCM can be adopted to other regions of the globe,
but results will differ on which RCMs is used. A single RCM will most likely not provide
accurate results, and therefore one must keep to mind the above dynamical downscaling
assumptions.

Most intensive downscaling studies and projects use many RCMs to produce a multi-model
ensemble and further validate results against observations. This usually depends on the type of
project, the region and purpose of the project, which determines the method of RCMs used. For
example, the use of Eight RCMs, Five RCMs, Multi-RCM comparison, and dynamical and
statistical methods.

Statistical Downscaling: Refers to the establishment of empirical correlation between historical

large-scale atmospheric and local climate characteristics. After tying the two and validated,
future large scale atmospheric conditions projected by GCMs are used to predict future local
climate conditions. Large-scale GCM outputs are used as predictors, to get local variables or
predictands, which involves heterogeneous group of methods with varying sophistication and
applicability.

Statistical downscaling methods are computationally inexpensive in comparison to RCMs that

need complex modeling of physical processes, which is better than dynamical downscaling in
that respect. Plus, unlike RCMs, which provide downscaled projections at spatial scale of 20-

7
50km, statistical downscaling can provide a station-scale climate information, the best a small-
holder farmer can have and use.

Assumptions of statistical downscaling :

1. The statistical relationship between the predictor and the predicand does not change over
time .
2. The predictor carries the climate change signal.
3. There is a strong relationship between predictor and predicted and,
4. GCMs accurately simulate the predictor.

Assumptions I is known as the stationarity assumption and states that predictor and predictand
relationship remains stable. The 2nd assumes that large-scale variables represent the climate
system and captures any charges that may occur in the future. The 3 rd assumption implies that the
strength of the relationship determines its validity, and the 4 th relates the ability of GCM to
simulate climate variables observed in the past as well as in future evolution.

Statistical downscaling methods can be classified into three main categories:

1. Linear Methods: This establishes a linear or proportionate relationship between the

predictor and the predictand. This is usually applied in a single predictor-predictand pair
of spatial fields. All that is needed is a normal distribution of the predictor and predictand
values. It is therefore not suitable for distribution of daily rainfall where frequent small
amounts and heavy events make the distribution non-normal. This method is primarily
used for spatial downscaling.
2. Weather classification: This method relies on large-scale atmospheric states to predict
local variables. The stimulated GCMs future atmospheric states are matched with its most
similar historical atmosphere state. The selected atmospheric state then corresponds to a
value that can be replicated under future atmospheric states.
This method is good for downscaling non-normally distributed data such as daily rainfall.
However, a large amount of observed daily data, for example, 30-50 years data of a
region, is required in order to evaluate all possible weather conditions. This means that
computational demand is high due to large daily data analyzed and generated.

8
 3. Weather Generators: These statistical methods are used in temporal downscaling, needed
for some impact models that require local spatial data at a daily resolution, which GCMs
cannot provide. Weather generators provide sequences of daily values, and are data-
intensive, and are sensitive to erroneous or missing data in the calibration set. (Wilby et
al., 2009).

Table: Showing statistical downscaling category, method, predictor and predictand variables,
advantages, and disadvantages.

9
Source: Stardex,2005; Fowler et al., 2007; Wilby et al., 2009 and Daniels et al., 2012

Table: Showing advantages, disadvantages, outputs, requirements, and applications of dynamical

and statistical downscaling

Source:
Stardex,2005;
Fowler et al.,
2007; Wilby et
al., 2009 and
Daniels et al.,
2012

Conclusion

All said,
downscaling
climate change
projections to
the local level is
10
important to the small-holder farmers who are the most vulnerable to the impacts of climate
change. By providing region-specific and/or station-level climate information and adaptation
strategies, farmers can prepare well for and adapt to changing climate patterns. Some important
points that prompt downscaling are as follows.

1. Localized weather forecasts: Small-holder farmers rely heavily on seasonal weather

patterns for agricultural decision making. Downscaling global climate models (GCMs)
can go a long way in helping farmers plan for planting, irrigation, and harvesting times
more effectively (Thorton, 2014).
2. Crop Suitability Mapping: Downscaled climate data can be used to find suitable crops for
planting under changing climatic conditions. (Aggrawal, 2009) By identifying crops that
are resilient to water stress, temperature fluctuations, and pest outbreaks. Small-holder
farmers can diversify their choices of crops and adapt to climate variability. This crop
maps based on downscaled climate projections can guide farmers in making informed
crop selection decisions.
3. Agro-ecological Zoning: Agro ecological zones are areas with similar climatic conditions
that shows their ability to support rainfed agriculture. They are land resource mapping
units defined by climate, landform, soils and/or land cover. They are areas with
homogenous sets of climatic parameters and natural resource characteristics such as solar
radiation, rainfall, soil types, temperature, etc. Small-holder farmers can use this
information to select suitable crops, livestock breeds, and even farming practices that best
fit their specific zones. (Teixeira, 2019). Agroecological zoning can enhance agricultural
productivity and climate resilience at the local level.
4. Water Management Strategies: Downscaled climate projections can inform small-holder
farmers by predicting changes in rainfall patterns, water availability and variability, so as
to plan irrigation practices, water conservation techniques, and rain water harvesting
system to cope with water scarcity and even drought (O’ Brien, 2004) Access to localized
climate information is vital for water resources management on small farms, and this can
be gotten from RCMs.
5. Livestock Heat Stress Mitigation: In the face of climate change, rising temperatures are
imminent, and small holder livestock farmers are vulnerable to heat stress events caused
by this rising temperatures. Downscaled climate data can help identify hotspot where

11
 animals (and plants) are at risk, so as to take mitigation measures, such as shading,
ventilation, and water provision (Lobell, 2008). These targeted interventions can help
improve animal welfare and productivity in the glaring face of climate change.

The general benefits of downscaling global and/or regional climate information to local-level
scale are; but not exhausted by;

1. It allows farmers to better grasp how climate change will impact their own regions and
localities therefore, downscaling global climate models will provide long-term trends and
patterns, so that farmers may capture the uniqueness of their local climates and how it can
impact their crops, livestock, and livelihoods.
2. It allows farmers to adapt their present agricultural practices in response to climate
change conditions, by gaining information about changes in temperature, precipitation,
humidity, winds etc., at the local level, and to help them optimize their yields, reduce
risks, and improve resilience to climate change.
3. It allows planting and facilitating development and implementation of targeted adaptation
strategies for the small holder farmers, by identifying the specific climate risks and
vulnerabilities faced by particular farmers and a particular region. Downscaling can
provide the tailored priorities and needs for the local communities. This will ensure the
allocation of limited resources to the climatically challenged small-holder farmers.

12