1.1. Objective.

     The purpose of this project was to generate snowfall and snow depth statistics for several thousand non-airport stations in the National Weather Service (NWS) Cooperative (COOP) Network. Stations in the Lower 48 States and Alaska were considered. The Automated Surface Observing System (ASOS) instrumentation being installed at airport locations detects weather phenomena, including the occurrence of snow, using a standard observational methodology, however the ASOS automated instruments are not able to measure the amount of snowfall or snow on the ground (snow depth). This project established snow climatologies for COOP stations which could be used to support NWS real-time snow operations in the ASOS observation era. These snow climatologies also enable the National Oceanic and Atmospheric Administration (NOAA) to better respond to user requests for snow information for use in economic and engineering decision-making, and provide the Federal Emergency Management Agency with an objective basis for declaring federal snow disasters.

     This project analyzed daily snowfall and snow depth data from NCDC's TD-3200 Cooperative Summary of the Day data base. The digital period of record was examined. Daily maximum and minimum temperature and precipitation were used to quality control (QC) the snow data.

     Three levels of quality control were employed in order to obtain the best snow data possible. The first level involved using the ValHiDD edited TD-3200 values. The second level employed a number of internal consistency checks. The third level was an extremes check.

     During the 1990's, an automated quality control system called ValHiDD (Validation of Historical Daily Data) was applied to the entire TD-3200 data base to remove gross errors in daily maximum and minimum temperature, precipitation, snowfall, and snow depth. ValHiDD is a rules-based method for detecting and correcting discrepancies (due to digitizing errors and observer errors) in the TD-3200 data base. The checks employed by ValHiDD include a limits check, internal consistency checks, flatliner temperature check, precipitation/snowfall/snow depth (PSFSD) relationship check, temperature spike check, multiple rule-group failures check, and failed fix check ( Reek, et al., 1992).

Although the number of discrepancies uncovered and resolved by ValHiDD was small compared to the total number of data values examined, their removal/correction was important and ValHiDD significantly contributed to the improvement of the overall TD-3200 data base. However, two factors relevant to this project should be noted:

(1) In some PSFSD cases, ValHiDD could not identify which element should be corrected, so the values were flagged as suspect and not altered.

(2) The PSFSD relationship check assumed that all three elements were observed at the same hour. For most volunteer COOP observers this assumption holds. However, for airport stations, this is not the case: snow depth is observed at 7 a.m. local time, while daily snowfall and precipitation amount are reported as of midnight. This airport station observation time discrepancy complicated the PSFSD relationship check.

As in the ValHiDD discussion (see section 1.3.1), it is critical to these automated tests that the temperature, precipitation, snowfall, and snow depth observations be taken at the same hour. This is the case for most volunteer COOP observers. However, NWS and Federal Aviation Administration (FAA) airport stations observe snow depth at 7 a.m. local time, while the remaining elements are reported as of midnight. This airport station observation time discrepancy impacts checks (4)-(7) and (10) above, and can result in valid snow depth values being flagged as erroneous and being changed.

Snow depth varies widely in states with mountain topography. For example, the extremes for coastal stations in southern California would be considerably lower than the extremes for stations in the Sierra Nevada range. This made it difficult to establish an appropriate statewide snow depth extreme, so a standard snow depth extreme of 2000 inches was used for all stations in all 49 states. If a station's snow depth value exceeded 2000 inches, the value was set to missing and the occurrence was similarly tallied and saved.

Two sets of QC and inventory indicators were computed for each station, one for snowfall and one for snow depth. As a result, some stations will have output products and statistics for snowfall but not snow depth, some for snow depth but not for snowfall, and some for both snowfall and snow depth.

     Daily COOP data have been digitized operationally beginning in 1948. Over the years, pre-1948 data have been keyed on a special project case-by-case basis, and are more likely to have gaps of missing data. About 18 percent of the stations had data beginning before 1948, while less than one percent (0.36%) had data which ended before 1948. Consequently, the QC indicators were computed for the 1948- 1996 period (however, the entire data base was QC'd and analyzed for the computation of the statistics). The QC indicators include the following:
    (1) the number of non-missing daily values read;
    (2) the number of daily values that were flagged as suspect by the QC checks, including those flagged values that were corrected and those flagged values for which no corrective action was taken;
    (3) the number of daily values that failed the quality control checks and were set to missing; and
    (4) the percent of the non-missing daily values read that were flagged as suspect or set to missing.

    For the reasons noted in Section 1.4.1, the inventory indicators were computed over the period 1948-1996. They include the following:
    (1) number of years in the TD-3200 data base between the first and last years with data;
    (2) number of years in the TD-3200 data base having some data (at least one day);
    (3) number of months having complete data (no days missing), and percent of possible months having complete data;
    (4) number of usable daily values processed;
    (5) number of daily values missing, and percent of daily values missing; and
    (6) information concerning the number of breaks (or gaps) in the data record, where a break is defined as at least one month completely missing. The information included the number of breaks of different lengths, the total number of breaks (breaks with any number of months missing), and the length of the biggest break. For example, if a station had one break of three months length and two breaks of five months length, then it would have two breaks of different lengths, three breaks in total, and the biggest break would be five months long.

   Ob time information is available from only 1981- present. There are three pronounced peaks in a plot of ob time change: one at 1 hour, one at 9-10 hours, and one at 16-17 hours. One can safely assume that an ob time change of 1 or 2 hours will not introduce a significant inhomogeneity into the snow record. Ob time changes of 3-5 hours are rare. For these reasons, some tolerance was built into this indicator, as well, with an ob time change being counted only if it exceeded 2 hours.

(3) from the HCN file: number of years the station is in the file, the number of times the station changed location, and the number of times the station changed location divided by how long it is in the file.
      The discussion for the COOP location indicator applies to the HCN location indicator, except the HCN location information was available for the entire period of record.

(4) from the HCN file: the number of times the station's ob time changed, and the number of times the station's ob time changed divided by how long it is in the file.
      The discussion for the COOP ob time change indicator applies to the HCN ob time change indicator, except the HCN ob time information was available for the entire period of record.

(5) from the HCN file: the number of times the station's observer name changed, and the number of times the station's observer name changed divided by how long it is in the file.

(6) from the CLIM81 file: number of years the station is in the file, the number of times the station's ob time changed, and the number of times the station's ob time changed divided by how long it is in the file.
      The discussion for the COOP ob time change indicator applies to the CLIM81 ob time change indicator, except the CLIM81 ob time information was available for the CLIM81 file's period of record.

     The following QC, inventory, and station metadata indicators were used to assess the quality of each station. Examination of frequency distribution charts of these indicators did not provide meaningful guidance in determination of cutoff criteria. Therefore, the specific criteria chosen were selected in order to maximize both the quality of the station dataset and the number of stations included in the dataset. The data and metadata indicators for all stations are included in the metadata files, in the event the user wishes to apply different criteria.
     In order to be included in the project's final station list (for snowfall and/or snow depth), the station had to meet the following requirements:

(1) have at least 15 years of non-missing data for each of the 12 months (January-December) for selected climatic elements (number of days with snowfall, monthly total snowfall amount, greatest daily snowfall amount, number of days with snow depth, and daily snow depth amount);

(2) have at least 15 years of non-missing data for each of the 365 days of the year for selected climatic elements (daily snowfall amount and daily snow depth amount);

(3) have at least 70% of the months from the data period with complete data (no days missing);

(4) have 33 or fewer breaks per 100 years;

(5) have no more than 25% of the daily values missing out of the total number of (daily values missing plus daily values with data);

(6) have 3 or fewer ob time changes, based on the COOP metadata file;

(7) have 10 or fewer location changes per 100 years, based on the COOP metadata file (it should be noted that latitude, longitude, and/or elevation may have changed due to the switch from manually-based surveys to satellite-based surveys when, in fact, the station did not move); and

(8) have 10 or fewer ob time changes per 100 years, based on the CLIM81 metadata file.

      Stations which met these criteria were then examined for station type. NWS offices and NWS and FAA airport stations were deleted from the snow depth list, due to QC considerations as discussed in sections 1.3.1 and 1.3.2.

     A suite of statistics (mean, median, first and third quartiles, extremes [both amounts and dates of occurrence], and frequencies/probabilities) was generated for several climatic parameters. The specific statistics that were computed vary with parameter, but the number of years with non-missing data (NYRS) was computed for each parameter. The beginning and ending years and NYRS information are crucial for any inter-station or inter-seasonal comparisons the user may wish to make.

     A daily climatology and a monthly/seasonal climatology were created for each station. The statistics for the daily climatology were generated for each day from the years of data available for the day. The statistics for the monthly/seasonal climatology were generated from year-month or year-season sequential values.

     The climatic parameters include the following:

(1) number of days (1-day periods, 2-day periods, and 3-day periods) with daily snowfall amount equal to zero or a trace;

(2) number of days (1-day periods, 2-day periods, and 3-day periods) with daily snowfall amount greater than or equal to several thresholds (0.1, 1.0, 2.0, 5.0, 10.0, 12.0, 18.0, 24.0, and 36.0 inches);

(3) monthly and seasonal total snowfall amount;

(4) number of consecutive days with daily snowfall amount greater than or equal to several thresholds (0.1, 1.0, 2.0, and 5.0 inches);

(5) dates of first and last occurrence of daily snowfall amount greater than or equal to several thresholds (1.0, 4.0, and 6.0 inches);

(6) daily snowfall amount, both with all days examined (whether they had snowfall or not) and only days having snowfall examined;

(7) greatest multiple-day (2-, 3-, 4-, 5-, 6-, and 7-day) total snowfall amount (where snow fell on each day) in a month;

(8) greatest 2-day and 3-day total snowfall amount (whether it snowed each day or not) in a month;

(9) number of days with snow depth equal to zero or a trace;

(10) number of days with snow depth greater than or equal to several thresholds (1.0, 2.0, 5.0, and 10.0 inches);

(11) number of consecutive days with snow depth greater than or equal to various thresholds (1.0, 2.0, 5.0, 10.0, 12.0, 18.0, 24.0, and 36.0 inches);

(12) daily snow depth amount, both with all days examined (whether they had snow cover or not) and with only days with snow cover examined; and

(13) number of years (from which frequencies were derived) with snowfall or snow depth greater than or equal to several thresholds (0.1, 1.0, 2.0, 5.0, 10.0, and 12.0 inches) for each day of the year.

     The multiple-day extremes parameters (see group 7 in Section 1.5.1) are based on snow falling on each of the days in the time unit. If snow fell on some days but not all days, then that value was not included in the analysis for that multiple-day time unit, but it may qualify for inclusion in a shorter time unit. This could result in some longer time units having smaller extreme snowfall amounts than the corresponding shorter time units.
     The date (year, month, and/or day) of an extreme is the date of the most recent occurrence (except for statistics G1-G0). The date listed for multiple-day parameters is the last day of the multiple-day period.

     Monthly statistics (for January through December) were computed based on the days in the month under consideration. Seasonal statistics were computed for winter, spring, summer, autumn, annual, and snow season, with the seasons corresponding to the following months, respectively: December-February, March-May, June-August, September- November, January-December, and August-July. The seasonal statistics are not based on the monthly statistics; they were computed from the daily values corresponding to each season in each year of the record. (For example, the mean winter statistics are not the average or total of the December, January, and February mean statistics; they are based on the sequential winter periods in the data record.) For the first and last occurrence of snowfall (group 5 in Section 1.5.1), the (incomplete) years at the beginning and end of the data period were included in the analysis for the seasonal statistics. For these reasons, the seasonal statistics may not agree with the corresponding monthly statistics.

     For the first and last occurrence of snowfall (group 5 in Section 1.5.1), if no snow occurred, then there was no data from which to compute the dates and the first and last years of data will be coded as "missing" (-99). For these elements, the NYRS statistics refers to the number of years with nonzero data, which is somewhat broader than "the number of years with non-missing data."

     Likewise for the daily snowfall (snow depth) amount, where only days having snowfall (snow depth) were examined (groups 6 and 12 in Section 1.5.1): if no snow occurred, then there was no data from which to compute a value. Consequently, the NYRS statistics refers to the number of years with nonzero data, which is somewhat broader than "the number of years with non-missing data."

     The greatest number of consecutive days with snow depth parameter (element 71) was computed for just the August-July snow season. It was not computed for the other seasons or the individual months.

     The impact of missing data varies, depending on the element and statistic computed. Total snowfall amount had no tolerance for missing data. If even one day was missing in a month or season, the total snowfall could not be computed for that year's month or season. Consequently, the number of years with non-missing data will vary with month and season. The six seasons (especially annual and August- July) have a greater chance of experiencing missing data and, generally, will have fewer years with non-missing data when compared to the individual months.

     The median daily value for a month had no tolerance for missing data. All days in a month had to have data in order for a median daily value to be computed for that year-month.

     The number of days with snowfall or snow depth parameters had no tolerance for missing data. Data for leap days (February 29) were included in the analysis. Due to this fact and due to rounding error, the sum of the values for the equal zero, equal trace, and greater than or equal to 0.1 inch (for snowfall, 1.0 inch for snow depth) thresholds may not exactly equal the maximum possible number of days in the month or season. This will be especially noticeable for the number of days with 2-day and 3-day snowfall parameters.

     The consecutive days with snowfall or snow depth, or "runs" parameters, had no tolerance for missing data for each specific threshold. A run of consecutive days for a given threshold was delineated by days (immediately before and after the run) which had values less than the run's threshold value. Consequently, a run had to have no missing days during the run and on the days immediately before the run started and after the run ended in order to be included in the analysis. If a nonzero day outside of this range was missing, however, then the runs for the lower thresholds would be affected and would be treated as missing in the sequential data for that year. This could result in statistics for runs with higher thresholds being larger than the corresponding statistics for lower thresholds.

     The daily extreme, multiple-day extreme, and date of occurrence parameters had a greater tolerance for missing data. Data were analyzed even if a month had up to 5 days missing. This could result in apparent discrepancies between these and other parameters.

     The greatest 2-day and 3-day total snowfall amount, whether it snowed each day or not (group 8 in Section 1.5.1), could tolerate up to 5 missing days per month. However, missing days within a 2- or 3-day period were excluded from the analysis. (For example, if a 2-day period had snow, but the day before and the day after this 2-day period were missing, then the snowfall total would be included in the 2-day analysis but not in the 3-day analysis.) Missing days might, in some cases, result in extreme 2-day snowfall totals being greater than the corresponding extreme 3-day snowfall totals.

     A complicated tolerance for missing data was built into the probability of receiving measurable snowfall parameter (statistic PR). The data in a month (or season) were examined to determine if it snowed in a given year-month (or year-season). If even one day in a month (or season) had measurable snowfall, then that year was counted in the computations, regardless of how many days were missing. However, to determine if no snow occurred in a month (or season), the month (season) had to have no days missing. The probability for each month (season) was computed by summing the number of years with one or more days of measurable snowfall, then dividing by the number of years that qualified (i.e., where it could be determined that it did or did not snow).

 2.1. Objective.

     The purpose of this project was to prepare snowfall extremes and return period statistics for official weather stations across the contiguous United States and Alaska, which the Federal Emergency Management Agency (FEMA) may use as an aid in making disaster declarations for record or near-record snowstorms. This work was performed as an adjunct to the Production of Snowfall and Snow Depth Climatologies for NWS Cooperative Observer Sites project.

     This project analyzed daily snowfall data from NCDC's TD-3200 Cooperative Summary of the Day data base. The digital period of record was examined. Daily maximum and minimum temperature and precipitation were used to quality control (QC) the snow data.

     Three levels of quality control were employed in order to obtain the best snow data possible. The first level involved using the ValHiDD edited TD-3200 values. The second level employed a number of internal consistency checks. The third level was an extremes check.

     Same as Section 1.3.1 of the Production of Snowfall and Snow Depth Climatologies for NWS Cooperative Observer Sites project.

     This level of QC included additional inter-element checks beyond those performed by ValHiDD. Snowfall values which failed the internal consistency checks were corrected (where possible) or set to missing. Temperature or precipitation values were not examined for accuracy at this level.

The second level QC included the following checks. The following abbreviations are used here: TMIN = minimum temperature (deg. F), TMAX = maximum temperature (deg. F),
P = precipitation (inches), SF = snowfall (inches), and SD = snow depth (inches).

(1) Factor of 10 error for SF: if P >= 0.01 and SF >= 1.0 and the ratio, SF/P, was greater than 80.0, then the SF value was corrected by dividing by 10. The corrected SF value was similarly checked and set to missing if the new SF/P ratio was greater than 50.

(2) Hail check: nonzero SF values were set to zero if TMIN >= 50, or TMIN >= 40 with
TMAX >= TMIN + 20.

(3) Nonzero SF values were set to missing if: SF > 0.4 but P = 0.

     The snowfall values which were corrected or set to missing (by the above checks) were tallied and the counts were saved to a metadata file to be used later in a station quality assessment step.

     The following additional checks were made. Values failing these checks were not changed, but the number of flagged values was similarly saved to a metadata file.

(4) Questionable SF values (the SF/P ratios were unusual):
       (I) 1 > SF < 3, and SF > 50*P; or
       (II) 3 <= SF <= 6, and SF > 40*P, and TMAX > 24; or
       (III)SF > 6, and SF > 20*P, and TMAX > 24; or
       (IV) SF > 6, and SF > 30*P, and TMAX < 25.

     The daily snowfall values were compared, on a state-by- state basis, to known statewide 24-hour snowfall extremes. The known extremes published by weather historian David Ludlum ( Ludlum, 1982) were, for most states, multiplied by an acceptability factor (1.4) in order to account for new daily extremes that may have been set since his book was published, and to account for the difference in time frame (a moving 24-hour time frame versus daily values taken at a fixed ob time). Special subjective estimates were used for Colorado, Florida, and New York.

     These adjusted statewide extremes were used in the snowfall extremes check. If a station's daily snowfall value exceeded the corresponding statewide extreme, the value was set to missing and the occurrence was tallied. The counts were saved to a metadata file to be used later in a station quality assessment step.

     FEMA needed snow statistics for as many stations as possible, so all of the "current" stations were selected. A station was considered to be "current" if it had snowfall data in the TD-3200 data base in 1996, or if it was listed as a currently-open station in the Cooperative Station History File. Some stations were listed as currently open yet did not have current snowfall data in the data base. These were included in the product sent to FEMA. The NCDC archive product, however, includes these stations as well as non-current stations that were included in the station list for the Production of Snowfall and Snow Depth Climatologies for NWS Cooperative Observer Sites project.

     As noted in the Analysis Procedures section for the Production of Snowfall and Snow Depth Climatologies for NWS Cooperative Observer Sites project ( Section 1.4), the properties of snow make it difficult to accurately and consistently measure snowfall. The QC, inventory, and metadata quality assessment indicators computed by that project for all stations should be consulted before using the return period statistics of the FEMA project.

     Data from the digital period of record were analyzed. The output for this project includes observed extreme snowfall values and the extreme snowfall values that correspond to four return periods. These values were computed as follows:

(1) Four values (corresponding to four time units) were determined for each year of the data period: the greatest 1-day, greatest 2-day, and greatest 3-day snowfall amounts, and the August-July total snowfall amount.

(2) Each time unit was analyzed separately. For example, 1-day snowfall might have had 35 values (35 extreme values, one for each of 35 years), 2-day snowfall might have had 30 values, 3-day snowfall might have had 26 values, and August-July total snowfall might have had 21 values.

(3) The highest of these values was selected as the observed maximum snowfall value.

(4) These extreme values were analyzed using the Generalized Extreme-Value statistical distribution estimated using the method of L-moments and L-moment ratios described by Hosking and Wallis (1997). This analysis method can be used to compute the extreme snowfall values corresponding to any desired return period (i.e., probability level). Extreme snowfall values corresponding to the 10-year, 25-year, 50-year, and 100-year return periods were computed for each of the four time units.

     It should be noted that a statistical distribution can be determined only from nonzero values. If too many values are zero (which will happen, for example, in warm climate regions such as the Gulf Coast states, southern New Mexico, southern Arizona, and coastal and southern California, where it rarely snows), then a statistical distribution cannot be determined and return period statistics cannot be computed. Hosking and Wallis note that at least 20 nonzero values are needed in order to determine the statistical distribution, but a study by Guttman (1994) indicates that at least 30 nonzero values are needed for stable return period statistics.

     In this study, return period statistics were computed if at least 20 nonzero values were available, in order to generate return period statistics for as many stations as possible. The number of years with nonzero data are included in the output to enable the user to decide if they want to use a particular station's values. If the number of years with nonzero data is 20 or more but less than 30, then the user should exercise caution when using the return period statistics for that time unit.

     Some time units may have had 20 or more years with nonzero data, while other time units for the same station may have had fewer years. In these cases, only some of the return period statistics were computed; the others have a "not available" code.

     Guidance from the World Meteorological Organization (WMO) was used, where available, for determining the values for each year. The August-July total seasonal snowfall was computed by adding the total monthly snowfall amounts from the corresponding months for each August-July "year." If any month was missing, the August-July value for that "year" was missing. The total monthly snowfall amounts were computed by summing the daily snowfall amounts in the month. If any daily snowfall value was missing, the monthly total was missing. In this way, the August-July total seasonal snowfall has no tolerance for missing data.

     For the 1-day, 2-day, and 3-day snowfall extremes, the data for a month was discarded if more than five days were missing. If five or fewer days were missing, then the highest value for that month was used for the month. The extreme value for the August-July snow "year" was the highest value of the twelve available corresponding months. It may be possible, for some locations that experience only a few days of snowfall each year, to have a year with no snowfall when, in fact, snow did fall but the data was discarded because it failed a QC test.

     As noted above, the August-July total seasonal snowfall has no tolerance for missing data. The 1-day, 2-day, and 3-day snowfall extremes are more tolerant of missing data. For this reason, the August-July total seasonal snowfall will always have the same or fewer number of years with non- missing data than the 1-day, 2-day, and 3-day extremes. It may be possible for the 1-day, 2-day, or 3-day extreme value(s) to be greater than the August-July extreme seasonal total if the corresponding year(s) for the August-July total value was missing.

     The 1-day extreme snowfall value is the greatest single daily snowfall amount. The 2-day extreme snowfall value is the greatest two-day snowfall amount, where data were available for both days. If heavy snow fell on one day but the day before and the day after were both missing, then that snowfall amount could not be included in a two-day total or a three-day total. In this way, it may be possible for the 1-day extreme snowfall value to be larger than the 2-day or 3-day extreme snowfall value. Likewise for 3-day snowfall. The 3-day extreme snowfall value is the greatest three-day snowfall amount, where data were available for each of the three days. If heavy snow fell on one or two days, but the day before and the day after this one-day or two-day period were missing, then that snowfall amount could not be included in a 3-day total. In this way, it may be possible for the 1-day or 2-day extreme snowfall value to be larger than the 3-day extreme snowfall value.

     The users of these snow extremes and return period statistics should use considerable caution if they compare the values from one station to those of another station. The stations may not be directly comparable to one another, due to several reasons:

(1) The properties of snow make it difficult to accurately and consistently measure snowfall. As noted by Doesken and Judson (1997), snow often melts as it lands or as it lies on the ground, snow settles as it lies on the ground, and snow is easily blown and redistributed. These properties can be affected by location, time of day the observations are taken, and how often they are measured.

(2) The synoptic weather patterns that generate snow can result in snowfall amounts that vary greatly over small distances (snow bands).

(3) Local topography can have a major effect. Snowfall amounts can vary greatly depending on elevation and on slope aspect, steepness, and orientation (especially with regard to the prevailing wind patterns and the wind patterns associated with any given storm).

(4) The data period is an important factor. Ideally, the same data period (with no missing data) would be desired for all stations if any inter-station comparisons were to be made. In reality, the stations have varying data periods with differing amounts of missing data.

(5) The results of a statistical analysis partly depend on how much data are analyzed (sample size). A bigger sample size (60 or 70 years of nonzero data) would provide more stable results for this type of analysis ( Guttman, 1994). Unfortunately, this amount of data was not available. The preferred minimum sample size is 20 to 30 years of nonzero data, but the user should exercise caution when using return period statistics if the number of years with nonzero data is 20 or more but less than 30.

(6) Even if two stations have the same number of years with nonzero data, the history of snowfall at a location can affect the shape of the statistical distribution, which determines the snowfall amounts corresponding to the selected return periods.

Doesken, N.J. and A. Judson, 1997: A Guide to the Science, Climatology, and Measurement of Snow in the United States, Second Edition, Colorado State University Department of Atmospheric Science: Fort Collins.

Easterling, D.R., T.R. Karl, E.H. Mason, P.Y. Hughes, and D.P. Bowman (R.C. Daniels and T.A. Boden, editors), 1993: United States Historical Climatology Network (U.S. HCN): Monthly Temperature and Precipitation Data. Carbon Dioxide Information Analysis Center, Environmental Sciences Division, Publication No. 4500, ORNL/CDIAC-87, NDP-019/R3, Oak Ridge, TN.

Guttman, N.B., 1994: "On the sensitivity of sample L-Moments to sample size." Journal of Climate, vol. 7, pp. 1026-1029.

Ludlum, D.M., 1982: The American Weather Book, Houghton Mifflin Co.: Boston.

Hosking, J.R.M. and J.R. Wallis, 1997: Regional Frequency Analysis: An Approach Based on L-Moments, Cambridge University Press

National Weather Service, 1972: National Weather Service Observing Handbook No. 2: Substation Observations, First Edition, Revised December 1972 (Supersedes Circular B), U.S. Dept. of Commerce, National Oceanic and Atmospheric Administration, Silver Spring, MD.

Reek, T., S.R. Doty, and T.W. Owen, 1992: "A deterministic approach to the validation of historical daily temperature and precipitation data from the Cooperative Network." Bulletin of the American Meteorological Society, vol. 73, pp. 753-762.

     This project was supported by ESDIM grant NWS-97-04 and FEMA grant EMF-2000-IA-0039. The project leaders would like to specially thank Dr. David Robinson of Rutgers University, Mr. Nolan Doesken of the Colorado Climate Center, and Mr. Grant Goodge of Asheville, NC for their valuable input and suggestions, and Ms. Nina Stroumentova for valuable programming support.