Using low frequency Predictors for Downscaling to mid-latitude precipitation: two case studies in mountainous regions

In this paper temporal and spatial scale issues associated with statistical downscaling for precipitation under climate change scenarios are examined in two areas: the Animus River Basin in southwestern Colorado; and the Catskill Mountains in southeastern New York State. Both are important water resource regions. In recent years different methods for overcoming the spatial mismatch between GCM output and regions of interest, such as basins associated with water resources, have been developed. This is especially important for precipitation, which is more difficult to model than temperature due partly to its high spatial variability. Downscaling methods fall into two general categories: dynamic, which includes nested modeling experiments; and statistical. Statistical downscaling methods involve the empirical identification of relationships between free atmosphere variables and the local variable of interest.

Here, using a multiple linear regression method, downscaling models are compared that use independent variables smoothed at different time scales from daily to monthly, and smoothed at different spatial scales. Independent variables are derived from the NCEP / NCAR reanalysis; dependent variables are basin-mean precipitation derived from station observations. Comparison of models is based on monthly mean rather than higher frequency results because these are considered more relevant for climate change studies.

Results indicate that in many cases lower frequency, such as monthly, input provides results comparable or superior to higher frequency input. In addition, it is often true that information averaged over large regions (i.e. many grid boxes) can provide results comparable to information from the single grid box above the basins of interest. These results are important because (1) it is generally true that GCM results are more reliable at larger scales; and (2) monthly mean observations are more readily available and reliable than daily observations at many locations. Thus, in cases when the use of information derived from larger spatial and temporal scales is appropriate, downscaling models can be applied in a wider variety of locations, and we can have more confidence in results from downscaling models applied to GCM results.