5. Comparison of model results with observed trends

In light of the expanded scope of the simulations shown in this present study, it is also of interest to revisit and expand upon the earlier observational data analysis provided by Pielke et al. (1999). In the previous work, time series of July-August accumulated rainfall and mean shelter-level temperature observations from Everglades City, Belle Glade, and Fort Lauderdale, Florida, were presented. Here, data from these and several more stations (see Fig. 24 for all station locations) were compiled to provide long-term, regional-average time series. Furthermore, the July-August regional-mean time series of daily maximum and minimum temperature are provided in lieu of the single mean observed value shown in the previous study.

Fig. 24. Locations of observation stations used to compile regional-average long-term time series. [larger image]

Data for Arcadia, Belle Glade, Everglades City, and Fort Lauderdale, Florida, through 2000 were obtained from the National Oceanic and Atmospheric Administration (NOAA) National Climatic Data Center (NCDC) United States Historical Climatology Network (USHCN) Serial 2000 Temperature and Precipitation Dataset. The analysis of rainfall data from these stations was based on the unadjusted area-edited (i.e., original or "raw") USHCN monthly total precipitation data. These totals were screened by NCDC to flag outliers, defined as three standard deviations beyond the mean for the period of record. Temperature data were subjected to NCDC USHCN 2000 quality assurance procedures. Specifically, these data are from the "area-edited, time of observation" data that have been adjusted for maximum and minimum system bias and station moves, with estimated values for missing and outlier data. However, these data were not subjected to the NCDC adjustment for urban heat island effects. The NCDC data for all July and August periods were nearly complete, with the few instances of missing data estimated from nearby station records. The Florida Climate Center provided the data for the remaining stations shown in Fig. 24, with the exception of rainfall data for HGS1, which was provided by the South Florida Water Management District.

Several limitations should be considered when comparing this observational data with the model results. First, the observations are point specific, whereas the model values represent an average over the finite area covered by the corresponding 10 km X 10 km grid cell. Thus, it is possible that an observation at a particular point, otherwise free of error, will differ significantly from the corresponding model grid value, which itself could accurately portray the average over the grid cell area. In order to minimize this source of uncertainty, observational data were spatially averaged. The resulting composite time series were then used for comparison with model domain-average data. Second, instrument platforms for a given station designation were periodically relocated (often by as much as several kilometers) during the long-term period of interest. It is plausible that such relocations could have a marked impact on the observed trend for a particular station. Third, the number of stations with long-term records is rather limited, thereby constraining the sample size used to construct long-term time series of regional average data. Finally, observations of all three variables of interest were not always available from a particular station for all times during the long-term period of interest.

It must also be recognized that long-term time series of July-August observations and model simulations for only three July-August periods of interest that employ two land-cover datasets that were constructed to represent the long-term change do not constitute two otherwise identical statistical samples. The valid times for the two different land-cover scenarios used within a given pair of simulations correspond roughly to the end points of the long-term observational time series provided, so different results within a pair of simulations may provide physical insight into the possible impact of long-term land-cover change regional climate trends. However, a statistically rigorous comparison with corresponding observations would require data from model simulations for every July-August period used to construct the observational time series. Furthermore, each of these simulations would need to employ a land-cover database valid for each of the individual July-August periods in the time series. Such datasets are not available, rendering this task impossible at this time.

Fig. 25. Regional-average time series of accumulated convective rainfall (cm) from 1924 to 2000, with corresponding trend based on linear regression of all Jul-Aug regional average amounts. The vertical bars overlain on the raw time series indicate the value of the standard error of the Jul-Aug regional mean. [larger image]

The time series of the regional-average July-August rainfall for the period 1924-2000 is shown in Fig. 25 (insufficient data prevented a start date prior to 1924), along with a trend based on a linear regression that incorporates all years in the sample. The linear trend has a slope of -0.064, with a total decrease of 5 cm, or 12%, over the period of record. This is consistent with the percentage decrease of the domain-average convective precipitation from the model results. The standard error bars overlain on the raw time series indicate an appreciable spread among the individual station totals for a given July-August period, but the error value is typically less than 25% of the regional mean for its period. Eight out of ten of the time series for the individual observation locations (not shown here) used to construct the regional-mean precipitation time series exhibited a decreasing trend in July-August precipitation during their respective observational period of record. Station spacing and location, along with the considerations discussed above, preclude a detailed comparison of these individual trends with the spatial character of differences within land-cover pairs of simulations. In particular, the lack of long-term records for stations in the heart of the Kissimmee River valley prohibits an analysis of whether this area has actually seen an increase in rainfall. Prior to land-cover conversion, these areas were inaccessible wetlands and largely devoid of meteorological observation sites.

Fig. 26. Same as Fig. 25, except for daily (a) max and (b) min shelter-level temperature (oC). [click on each of the graphs above to view a larger version]

Lack of available long-term records for temperature observations at all the sites shown in Fig. 24 prohibited a start date prior to 1948 for the regional-mean time series of maximum and minimum temperature (Fig. 26). However, clear trends emerge in the available data, as the time series for both the July-August daily maximum and minimum temperature indicate warming. The trend from linear regression has a slope of 0.011^o (0.009^o) yr^-1, with a magnitude of 0.57^oC (0.46^oC) for the maximum (minimum) temperature increase over the period 1948-2000. The maximum temperature trend is in reasonable agreement with the model domain-average increase of 0.31^oC (Fig. 13) that occurs when the pre-1900 dataset is replaced with 1993 land use in the simulations. Recall, however, that the change in model minimum temperature indicated slight cooling (-0.26^oC for the grid average; Fig. 14). This discrepancy could result from a number of factors, including model error or inadequate observational sampling of those areas with the greatest cooling in the model that are in the heart of the poorly sampled Kissimmee River valley. In the case of model minimum temperature, the vertical resolution in the lower levels and the parameterization of stable boundary layer processes could result in inadequate representation of shallow nocturnal circulations. In addition, the subgrid-scale variability of minimum temperature is potentially greater than that for the daytime maximum, because the latter typically is more spatially homogenized when the boundary layer is more mixed. Furthermore, shelter-level temperatures are not predicted explicitly by the model but are diagnosed using similarity theory (Monin and Obukhov 1954). Numerous factors in the similarity theory framework could introduce error in the estimates, and this error may not be consistent for stable versus unstable (i.e., daytime maximum and nighttime minimum) conditions. Subtle differences between the elevation of the observing platform and the model grid cell effective elevation may also result in diagnosis of a temperature value that is not at the exact elevation of the observation. This error can be large when steep, surface-based thermal inversions are present, such as the time when the daily minimum temperature is observed. It must also be recognized that the observational time series could be indicative of a regional or larger-scale trend in nocturnal warming that is indeed real, and that it could be the result of factors independent of those addressed in this study.

Go back to Sensitivity Tests | Go ahead to Concluding Remarks