Uncertainty is inherent in SNAP’s work
Data analysis and interpretations are always uncertain, and must be understood in order to effectively and appropriately use our products. Scenario planning (allowing for more than one possible future) allows for greater flexibility in the face of high uncertainty. We also try to reduce uncertainty in these areas:
Interpolation, gridding, and downscaling of historical and projected data are uncertain
Differences between CRU and PRISM data can be considered a proxy for uncertainty in downscaling. PRISM generates gridded estimates using point data, a DEM, and other spatial data. CRU data uses different algorithms, and does not use data on slope, aspect, or proximity to coastlines. Generally, PRISM better captures fine-scale landscape climate variability. Caveats:
- SNAP historical raw climate data include weather station temperature and precipitation data every month of every year 1900–2006. Data are interpolated to a relatively coarse-scale grid using algorithms from CRU, and then further downscaled to a finer grid by SNAP using PRISM
- SNAP projected raw climate data include GCM temperature and precipitation data for every month of every year 1980–2099. Data are downscaled by interpolation between large scale grid cells (splining) followed by PRISM downscaling
- Interpolation, gridding and downscaling are challenging and imperfect regardless of method
- Climate stations are sparse in the far north
- Precipitation can vary enormously over very small areas and time frames
- Learn more about our data sources
Natural variability in historical and projected data creates uncertainty
Normal changes in weather patterns can obscure trends such as a warming climate. GCM outputs simulate this natural variability, but variations cannot be expected to match actual swings. Precipitation varies more across time and space and is thus more uncertain than temperature. Caveats:
- Averaging across all 5 models can reduce ups and downs built into the models
- Averaging across years (decadal averages) can reduce uncertainty due to natural variability
- Each method reduces the ability to examine extreme events
General Circulation Models (GCMs) include uncertainty
The 5 GCMs used by SNAP have been tested for accuracy in the North. Variation between models can be used as a proxy for uncertainty in GCM algorithms. SNAP’s model validation study shows uncertainty by region, model, and data type based on comparisons between model results and actual station data. Solar radiation is a known quantity. Future levels of greenhouse gases are uncertain, but accounted for by various emissions scenarios. Caveats:
- Interactions modeled in GCMs include thresholds (tipping points) such as ocean currents shifting or shutting down. However, oceanic and atmospheric circulation are extremely hard to predict and model despite SNAP’s use of the best GCMs.
- GCMs don’t fully account for short-term phenomena such as the PDO, which can affect Alaska’s climate over years or decades
- Averaging across all 5 models (using the composite model) can reduce potential bias — but also the ability to examine extreme events
- Learn more about 5 Global Climate Models chosen by SNAP
Linking models is useful but can increase uncertainty
Approaching the same question using multiple linked models can serve as a form of validation. We have "ground-truthed" linked models using historical data in all ALFRESCO runs as a means of calibration. Caveats:
- SNAP products that link our raw monthly climate data to other models must be interpreted in the context of the combined uncertainty of the raw data and the models to which these data are linked.
Uncertainty in current ecosystem modeling efforts by SNAP and collaborators
The ALFRESCO model uses SNAP input to project fire on the landscape and is well calibrated to match historical climate conditions to historical fire records.
Projections are inherently uncertain because they depend on assumptions and estimates regarding the frequency and location of fire starts as well as the calculated relationship between climate, forest age and type, and fire spread.
SNAP permafrost modeling has been performed in conjunction with experts at the Geophysical Institute Permafrost Lab (GIPL). Algorithms to determine the depth of active layer depend on calculations of the insulating properties of varying ground cover and soil types, as well as on climate variables.
Although GIPL researchers have used the best available data for all inputs, some datasets are incomplete.
Vegetation change: ecosystem modeling
SNAP has worked with multiple partners in the US and Canada to predict potential landscape shifts driven by climate change. Projections depend on links between vegetation and climate, as well as the ability of various species to shift across the landscape under either gradual or threshold-driven change.
Uncertainties stem from incomplete data on existing species ranges, behaviors, and dispersal, and incomplete data on the relationship between climate and habitat variables.
Uncertainty and interpolation references
Fleming MD, Chapin FS, Cramer W et al. 2000. Geographic patterns and dynamics of Alaskan climate interpolated from a sparse station record. Global Change Biology 6 (Suppl. 1): 49-58.
Sluiter R. 2009. Interpolation methods for climate data: a literature review. KNMI Internal Report, De Bilt, The Netherlands.