Drought and Model Consensus Reconstructing and Monitoring Drought

  • Slides: 1
Download presentation
Drought and Model Consensus: Reconstructing and Monitoring Drought in the US with Multiple Models

Drought and Model Consensus: Reconstructing and Monitoring Drought in the US with Multiple Models Theodore J. Bohn 1, Aihui Wang 2, and Dennis P. Lettenmaier 1 1 University of Washington, Seattle, Washington, USA; 2 Institute for Atmospheric Physics, Beijing, China UW Water Center Annual Review of Research, Seattle, WA, USA, Feb 14, 2008 Challenge: how to assess drought across the entire US in the face of sparse observations? Sparse observations Problem: how to assess errors? • Soil moisture sampling sites are sparse • Remote sensing techniques only scrape the surface (literally) • Historical records are limited in length • Without observations to constrain model estimates, how can we trust the model? Large-scale models fill in the gaps How do models compare in time? Soil Moisture Response Time • Define soil moisture response time = lag (months) at which autocorrelation of soil moisture (percentile) falls below 1/e • Measure of soil moisture “memory” Area-averaged soil moisture percentiles CLM 3. 5 soil moisture changes much more slowly than other models Solution: use multiple models • Mean of results tends to cancel random errors • Spread of results gives estimate of uncertainty • A model can transform our (relatively) dense network of meteorological observations into soil moisture estimates across the US Wide spread here = large uncertainty Large-scale Hydrological Models Example: VIC Typical Large-scale Hydrological Models Narrow spread here = more confidence • Break land surface into large (1/8 -degree), flat grid cells with uniform soil properties • Continental US = 3322 1/8 -degree grid cells • Most have vegetation layer, some allow multiple vegetation “tiles” • Multi-layer soil column • Input = daily or sub-daily meteorological data • Solve water, energy balances on sub-daily time step • Represent small-scale vegetation and soil dynamics with parameterizations All models agree that drier-than-normal conditions prevailed in the West and North during the 1930 s and that the South and Southeast experienced several dry periods. Models used in this study • VIC • CLM 3. 5 • NOAH • Catchment • Sacramento/Snow-17 (SAC) • Hybrid of CLM 3. 5 and VIC (CLM-VIC) This Study Retrospective Simulation: 1920 -2003 Response time (months) Models have different response times to climate variations Is this pattern realistic? • All models except Catchment, and both ensembles, exhibit much longer “memory” of soil moisture percentiles in the West than in the East • CLM 3. 5 has response times of several years in much of West • Ensembles have response times that are intermediate compared to range of model response times • Result: models and ensembles may tend to make dry/wet periods last longer in the West than in the East • Normally we think of drier soils, such as found in the West, as having shorter “memory” due to action of evaporation (“erasing” memory) than wetter soils, such as found in the East • This seems to contradict the pattern observed in our models • However, here we are computing the “memory” of percentiles (which have a different reference point each month) rather than “memory” of absolute soil moisture. Thus, we lose seasonality; this may help explain difference • Or, models may simply assume soils that are too deep in the West Domain: Continental US Examine Average Monthly Soil Moisture How to Compare/Combine Model Results? Simulated Average Monthly Soil Moisture (40. 25 N, 112. 25 W), 1920 -2003 Problem: Soil column has no common definition Assessing Model Agreement Severity-Area-Duration (SAD) Analysis Correlations among models Give indication of model agreement Average Model Correlation • Depth of soil column considered by a given model is arbitrary • Soil moisture storage capacities and dynamics differ widely among models VIC CLM 3. 5 SAC CLM-VIC NOAH Catchment ENS-0 ENS-1 Driest moisture level of CLM 3. 5 is wetter than wettest level of any other model Solution: “normalize” the soil moistures • Express as percentile of historical distribution (% of “normal”) Method 1 (Ensemble-0) Method 2 (Ensemble-1) • Express each model’s monthly soil moisture SM(y, m) as percentile of its distribution over entire period (1920 -2003) for that month, using Weibull plotting position • Average the percentiles to get ensemble average percentile SMens(y, m) • Note: this method de-emphasizes extreme values • Normalize each model’s monthly soil moisture SM*(y, m) = (SM(y, m) - SMmin(m)) / (SMmax(m) - SMmin(m)) where y = year; m = month • Average SM*(y, m) across all models to get ensemble average SMens*(y, m) • Express SMens*(y, m) as percentile of distribution over entire period (1920 -2003) for that month, using Weibull plotting position Western US Eastern US • Poor agreement • Larger uncertainty • Strong agreement • Smaller uncertainty Correlation and Response Times • In general, long response times (West) correspond to poor model agreement • Response times may affect uncertainty How do models compare in space? Example: droughts of 1930 s (dust bowl) and 1950 s SAD Analysis Most models (and ensembles) agree that: • Identify regions of contiguous dry grid cells • Categorize these regions by average moisture percentile, area, and duration • The 1930 s drought was characterized by large areas of high severity but short duration • The 1950 s drought was characterized by large areas of lower severity but longer duration • More recent droughts were characterized by smaller areas of high severity and moderate duration Monitoring Current Conditions Example: Recent drought in Southeast US Soil Moisture Percentiles wrt (1920 -2003) 2007 -Nov-01 2007 -Dec-01 Conclusions • An ensemble of multiple land surface models can reconstruct historical droughts across the continental US 2008 -Jan-01 • Ensemble average is more trustworthy than individual model • Cancels out disagreements among models • Multi-model ensembles can give us insight into uncertainty • Variation of uncertainty with location • Greater uncertainty in Western US, more confidence in Eastern US • Sources of uncertainty • Meteorology vs. model parameters • Long response times (West) correspond to poor model agreement • Model response times may affect uncertainty • Multi-model ensembles can be useful in realtime monitoring • Diversity of response times causes damped response to spurious real-time meteorological observations Ensemble-0 Little agreement among models over General agreement among models over South Colorado Plateau General agreement among models over Great Plains All models and both ensembles successfully capture the droughts of the 1930 s and 1950 s Ensemble average only exhibits extreme values where models exhibit strong agreement • Agreement with historic observations as to general areas affected • Agreement among models as to general shape and extent of affected area • Some disagreement as to details • Especially true of Ensemble-0 Real-time monitoring system • Drive models with real-time meteorological observations • Express daily soil moistures as percentiles of historical monthly distributions Benefits in a real-time monitoring system • Ensemble captures evolution of drought in SE US • Ensemble damps out disagreement in Western US • Different response times result in robustness against “spurious” meteorological observations