Most climatological datasets are beset with urban temperature influences that distort long-term trends. Using an hourly dataset of 41 urban and rural stations from the United States, discriminant functions were developed using diurnal temperature range indices based on temperature, dewpoint, and dewpoint depression that capture the differences between the two environments. Based on data for 1997–2006, diurnal temperature range and nighttime dewpoint depression range indices provide the best classification variables to statistically discriminate between urban and rural climates. Of the 41 stations, 93% were correctly classified by this technique in a cross-validation analysis. An additional discriminant analysis specific to coastal stations was needed because coastal climates were noted to be aberrant. Here, all stations tested were correctly classified by the procedure. Temporal trends in discriminant scores indicate periods of time during which urbanization was occurring or increasing. Instrumental and location changes were noted to affect both temperature and dewpoint series and therefore the classification. However, such discontinuities can potentially be adjusted and the homogenized data used with the classification technique. The use of this data-driven approach complements existing methods used to classify the urban character of stations, because it is objective, is applicable in the presatellite era, and can infer changes at higher temporal resolution.
Local climates are influenced by their surroundings, with urban heat islands (UHI) being one of the most well-known effects. Urbanization affects localized temperatures, but also moisture and wind patterns, with the effects of the cityscape being dependent on weather conditions (Arnfield 2003). Understanding the climatological influence of urbanization is important for identifying bias in temperature long-term records.
In most climatological studies the influence of the UHI on observations cannot be directly quantified, and therefore alternate methods are used. Population is often used as a proxy for city size because it has been established that UHI intensity is related to population (Oke 1973; Karl et al. 1988; Hua et al. 2008). Other urban proxies are indices of local economic activity (Schmidt 2009), vegetation indices (Gallo and Owen 1999), and satellite night-lights-derived urban/rural metadata (Peterson 2003). Peterson and Owen (2005) found the use of night lights, map information, and population metadata affected the number of stations that would be classified as urban. Over recent decades, such measures have been used to determine the degree to which urbanization and UHIs have affected temperature trends.
A UHI manifests from localized warming caused by the modified urban environment, relative to the surrounding countryside (Oke 1987). UHI intensity and development are frequently studied during clear and calm weather (Johnson 1985; Lu and Arya 1997; Parker 2006; Lee et al. 2009), because urban influences are strongest under these conditions (Morris et al. 2001). Changes in energy balance characteristics are also frequently used to elucidate the differences between urban and rural environments on energy fluxes (Oke 1982). Other methods of investigating UHI effects have been based on urban–rural transects, mesonets, combinations of transects and mesonets, and paired station observations (Morris et al. 2001).
Diversity in both urban and rural environments causes very dynamic spatial and temporal changes to the UHI. Rural temperature variability associated with changes in land use characteristics affects the magnitude of calculated UHI intensities (Hawkins et al. 2004). Stewart and Oke (2009) recently advocated a move away from the somewhat arbitrary binary urban/rural classification toward a more refined system based on “local climate zones.” However, the utility of such a classification method is limited when considering long-term changes. It has been shown that such long-term ecological change associated with population growth, infrastructure, and land use changes affects maximum and minimum urban–rural temperature differences (Brazel et al. 2000).
Cities have trended toward lower diurnal temperature range (DTR) than that in surrounding rural areas (Trenberth et al. 2007) because urban environments reduce nighttime cooling and increase minimum temperatures. Land use/land cover changes have been noted to affect DTRs within a 10-km radius of observation stations (Gallo et al. 1996). Temporally across all station environments in the United States, since the 1950s DTRs have shown a decreasing trend followed by a period of little change, associated with differential warming patterns in maximum and minimum temperatures (Vose et al. 2005). Factors that may have contributed to these changes in DTR are instrumental bias (Lin and Hubbard 2008), cloud cover/aerosol change (Zhou et al. 2009), and land use change (Kalnay and Cai 2003). Forster and Solomon (2003) found a weekend effect in DTR, which Jin et al. (2005) attribute to anthropogenic causes. In rural environments, irrigation affects latent heat fluxes and can reduce DTR by lowering maximum temperatures while having only a small effect on minimum temperatures (Mahmood et al. 2004; Lobell and Bonfils 2008). Annual temperature range in midlatitudinal coastal locations shows strong increases as distance from the sea increases over the first 15 km, after which temperature ranges show a slow increase as continentality increases (Sakakibara and Owa 2005). Marine advection also influences the UHI and can boost, limit, or reverse its effect (Runnalls and Oke 2000).
Urbanization has also been shown to affect atmospheric moisture content, but differences between urban and rural humidity measures are typically small and spatially complex (Oke 1987). Cities have been found to be moist at night and dry by day relative to the surrounding countryside (Ackerman 1971; Hage 1975; Oke 1987). Although a nighttime urban moisture excess (UME) is most prominent during summers under clear and calm conditions (Oke 1987), it can develop throughout the year (Kuttler et al. 2007). However, local moisture sources and energy balance influences can cause cities to have a UME during summer days (Unger 1999). In general, a positive correlation exists between UHI magnitude and UME (Mayer et al. 2003; Holmer and Eliasson 1999; Kuttler et al. 2007). The most intense UMEs develop after a UHI peak, indicating the role it plays in forming the UME (Kuttler et al. 2007). Relative humidity is a sensitive function of temperature and the urban–suburban/rural differences are strongly affected by the urban heat island (Unkašević et al. 2001).
Diurnal temperature range commonly has an inverse relationship with relative humidity; however, few studies explore changes in diurnal measures of atmospheric moisture. Near-surface dewpoint temperatures in the United States and Canada show an inverse relationship with temperatures and show some signatures of urbanization through relative reductions in dewpoints in early morning and late afternoon (Schwartzman et al. 1998). However, warm-season daytime increases in dewpoint relative to nighttime dewpoints are not an urban signal and could indicate a more widespread trend in dewpoints (Schwartzman et al. 1998). Anthropogenic activities and land use changes are linked to larger decreases in urban nighttime relative humidity in comparison with rural humidity levels in Taiwan, whereas daytime humidity levels show little change (Shiu et al. 2009).
A double wave occurs in rural absolute humidity, in which increases occur after sunrise because of evapotranspiration of moisture and before sunset associated with surface cooling causing stability and increasing moisture levels near the surface (Oke 1987). Mayer et al. (2003) also noted the presence of a double wave in vapor pressure differences generated from urban courtyards and park/green spaces. Dewfall has been implicated for drier atmospheric conditions in rural areas. Kuttler et al. (2007) showed the absence or delay of dewfall was the dominant effect in forming a UME. At urban sites, moisture is also higher at night because of warmer UHI temperatures, and an urban moisture minimum (relative to rural locations) occurs after sunrise because of the absence of evaporation of dewfall in urban areas (Oke 1987). Richards (2005) related dew to absolute humidity levels, but the nature of the surface was important, such as higher dewfall on rural grass and urban roofs. Rural landscapes and urban forests produce conditions that are favorable to dewfall, but urban environments generally inhibit dew (Ye et al. 2007).
This study aims to utilize the subtle differences in the diurnal patterns of temperature and atmospheric moisture between urban and rural locations as an objective means to classify stations as urban or rural. The rationale used expands upon a technique to detect microclimatic inhomogeneities using cooling ratios, developed by Runnalls and Oke (2006). Through the identification of microclimatic characteristics that differentiate between urban and rural climatic environments, a simple classification method is established. The method will supplement metadata regarding environmental changes near the observation sites, especially where records are sparse. It provides an alternative to previous more-subjective approaches that classify urban character based on population, land use map features, or artificial light prevalence. These approaches have limited utility, because they are merely proxies for urbanization and are not entrenched in any physical factors that relate to the climatic characteristics of observation sites. Furthermore, most urban classification methods do not directly relate to surface temperature and/or humidity regimes.
The new method developed offers advantages over other static approaches since urban–rural differences are based on objective, quantitative features of the station data. In addition, the classification method does not require detailed environmental metadata to identify periods in which urbanization occurred. These environmental changes potentially can be tracked, providing a temporal history of urbanization that can be applied to long-term temperature and dewpoint time series. With existing approaches, such histories are possible at best at coarse temporal scales.
2. Data and analyses
A set of 20 rural and 21 urban hourly reporting airport surface weather observations was selected to provide representative geographic and climatic coverage of the contiguous United States over the period 1950–2006 (Fig. 1). Metadata were obtained from the National Oceanic and Atmospheric Administration/National Environmental Satellite, Data, and Information Service/National Climatic Data Center (NOAA/NESDIS/NCDC) online station metadata archive database (Multi-Network Metadata System; http://mi3.ncdc.noaa.gov/mi3qry/login.cfm) for each site. Particular attention was paid to the most recent station information to check that observations were taken at typical airport environments, characterized by vegetated surfaces with adjacent paved runways. Specific microscale characteristics such as the features immediately adjacent to the instrumentation were not readily available. Therefore, subsequent analyses consider only the broad local-scale urban or rural effects in the vicinity of the observing site rather than the nuances of each particular observing site.
Stations were classified as urban or rural based on recent satellite images and aerial photography. A station was considered to be rural if the area within a 5-km radius of the airport was at least 75% natural (forested, crop land, and other natural land surfaces) and located at least 5 km from primary urban centers. Stations that did not meet the criteria for being rural were considered to be urban locations. The stations’ urban/rural classification, Weather Bureau Army Navy (WBAN) identification numbers, call signs, and current population are listed in Table 1.
Changes in nocturnal cooling rates during clear and calm weather that are similar to DTR have been shown to be useful in detecting microclimatic inhomogeneities in temperatures (Runnalls and Oke 2006). On this premise, and given the temporal features of the UHI and UME described in the literature, eight different indices were used as a basis for potentially differentiating urban from rural climates.
The following indices for screen-level temperature T, dewpoint DP, and their difference (dewpoint depression DPD) were calculated:
diurnal range midnight to midnight LST (TD),
diurnal range 0900 to 0900 LST (TD9),
nocturnal range 1500 to 0800 LST (Tn),
nighttime range 2100 to 0900 LST (TN),
temperature range forward, 6 h from the hour of each of the daily extremes (T6n; T6x), and
maximum 1-h warming/increasing DPD and cooling/decreasing DPD rates from 0300 to 1100 and 1200 and 2300, respectively (T1C; T1W).
Annual and seasonal averages of each index were calculated in degrees Celsius. The seasons were defined as winter (December–February), spring (March–May), summer (June–August), and autumn (September–November). Although clear and calm nighttime conditions are conducive to enhancing urban heat island effects, these conditions limit the number of days that can contribute to seasonal indices, particularly in cloudy and/or windy climates. Nonetheless, annual averages were also calculated for nighttime calm (wind speed < 2.2 m s−1) and clear and calm conditions (wind speed < 2.2 m s−1 and sky cover < 20%) separately because of the enhanced influence of the UHI under such conditions that has been reported in the literature.
b. Discriminant analysis
An iterative discriminant analysis procedure was used to assess which combinations of variables provided the best classification of urban and rural stations. A discriminant analysis was initially performed on the most recent 10-yr period of data (1997–2006). These data were selected because the systematic change to Automated Surface Observing System (ASOS) instruments occurred prior to this period and few subsequent siting changes exist at the 41 stations, thereby reducing the influence of inhomogeneities on the urban/rural classification. The predictors were assessed according to the maximum skill score obtained, using the Kuiper’s skill score (Wilks 2006). The Kuiper’s skill score measures skill from the proportion of stations that are correctly classified (hits) minus the proportion of incorrectly classified stations (misses). Although multiplicity problems are a concern because all combinations of the large-number variables (8 indices × 3 measured variables × 4 seasons) were assessed from only 41 stations, these issues were addressed through a randomized bootstrapping procedure. To assess the significance of maximum skill obtained, the urban and rural stations were randomly reclassified and assessed using the iterative discriminant analysis. This allowed the formulation of an empirical distribution of 1000 randomized maximum skill scores against which the statistical significance of the nonrandomized maximum skill could be assessed. A leave-one-out cross-validation method was applied to the selected discriminators to provide further verification that model skill would extend to independent stations not used in the development of the original model.
Further investigation of the temporal patterns of the best discriminant variables was then conducted using three long-term datasets from Baltimore, Maryland [Baltimore/Washington international airport (BWI)], Phoenix, Arizona (PHX), and North Platte, Nebraska (LBF). The environment surrounding the airports at BWI and PHX is urban while LBF is rural. Using the discriminant functions obtained from the 1997–2006 analysis, comparable discriminant scores were generated using 10-yr running periods beginning in 1950. Ten-year periods serve to smooth the variability inherent to nonoverlapping periods. Likewise the decadal time scales are typical of the interval over which urbanization generally occurs. Topographic maps produced since the 1950s and recent satellite images provided information on the general environment and urban-extent changes at these weather stations. Two of these sites (Baltimore and Phoenix) have been investigated for climatic change associated with increases in population and land surface modification (Brazel et al. 2000; Brazel and Heisler 2009). Maps of land use change produced by Brazel and Heisler (2009), in conjunction with metadata, were used to assess environmental influences on temporal fluctuations of discriminant classification and scores.
Population trends in Baltimore city show that growth ceased early in the twentieth century, but an expansion of built-up areas surrounding BWI (located 12 km southeast of the city) occurred between the 1970s and the 1990s. In contrast, Phoenix has shown rapid growth during the twentieth century and especially in the 1960s, with the area near the airport showing the greatest growth after the 1970s. Decreases in irrigated agricultural land as urban growth occurred also removed local moisture sources in Maricopa County near Phoenix (Balling and Brazel 1986). North Platte was selected as a third station because it is likely to have retained a similar rural environment throughout its record, with limited change in land use. This area has been irrigated from the late nineteenth century (Darton 1903) although, statewide, Nebraska irrigation has shown strong increases from the 1950s to 1970s.
Station data homogeneity is an important element of long-term climatological analysis, because site and instrument changes have been shown to affect temperatures (Della-Marta et al. 2004; Menne et al. 2009) and dewpoints (Robinson 2000; Brown and DeGaetano 2009). The homogeneity of Baltimore, Phoenix, and North Platte temperature, DTR, and dewpoint time series was assessed to determine the impact of nonclimatic inhomogeneities on the discriminant classification. Menne and Williams (2005) used multiple test statistics and composite reference series based on anomaly-weighted averages (Alexandersson and Moberg 1997), first-difference-weighted averages (Peterson and Easterling 1994), and a multiple linear regression method (Vincent 1998) to detect undocumented inhomogeneities in temperatures. Their methods were applied to annual averages generated from hourly datasets using reference series generated from the Cooperative Observer (COOP) network.
Temperature analyses followed the approach of Menne and Williams (2005). For dewpoint, the techniques of Brown and DeGaetano (2009) were used. Although data from the hourly WBAN dataset are recorded at a lower (hourly) resolution relative to the absolute minimum and maximum temperatures available from COOP data, the datasets are highly correlated and therefore are comparable. The homogeneity of this variable was therefore assessed using the method of Menne and Williams (2005).
3. Results and discussion
a. Urban/rural classification
Table 2 lists the highest skill scores achieved by each individual annual and seasonal discriminant variable. The highest discriminant skills of 56% for spring temperature cooling rates (T1C) and 46% for annual temperature cooling rates (T1C) illustrate that the temperature-based indices have stronger discriminant power in classifying urban and rural stations, when compared with dewpoint and dewpoint depression indices. To improve classification skill further, a second discriminant variable was added. All (4836) paired combinations of annual and seasonal predictors were assessed, and they showed that the relationship between DTR temperature indices and DTR of DPD indices is best able to classify urban and rural stations (Table 3). No maximum discriminant scores were produced from any of the dewpoint temperature indices. The addition of a third variable to the discriminant analysis was also assessed, but it failed to significantly improve on the predictability of the model relative to a randomly added third variable. An annual absolute humidity variable was also examined, but it did not improve upon the annual skills in Table 3.
The annual and seasonal distributions of maximum discriminant scores achieved through random resampling are shown in Fig. 2. The observed maximum discriminant scores of annual and seasonal two-variable classification combinations listed in Table 3 exceed the highest skill score given by the randomized resampling simulation. Thus, despite examining 4836 possible discriminant pairs, there is strong evidence that the two predictors selected can successfully differentiate between urban and rural stations. A leave-one-out cross validation was also performed to confirm that the discriminant pairs chosen retained their maximum skill scores and therefore that the results would be applicable to sites not used in their development.
Of the three variables that achieve an 85% maximum skill, annual averages of TD9 and DPDN provide the simplest indices for classifying urban and rural stations. Seasonal averages from spring and autumn perform no better than annual variables. There is also no benefit to examining diurnal range indices under specific weather conditions. Therefore the relationship between annual TD9 and DPDN was chosen for further analysis of temporal changes in urban classification. The dividing point m̂ between urban and rural stations is calculated from the average of the mean vectors for TD9 and DPDN, based on the 41-station analysis sample. Urban and rural stations are classified using the equation
where stations with m̂ < 3.281 are classified as urban and those with m̂ > 3.281 are classified as rural. The relationship between these two variables shows that, for a given TD9, urban stations have a larger DPDN than rural stations (Fig. 3).
Results from the discriminant analysis suggest that the microclimatic differences in diurnal temperature and moisture (dewpoint depression) characteristics described in the literature can be used to infer the urban or rural character of a climate station. The method also shows promise for determining periods of historical time during which stations undergo urbanization. For a given DPD range, the temperature range at urban sites is lower than for rural sites (Fig. 3). Lower DTRs are a typical feature of urban climates because of inhibited nocturnal cooling in urban areas. Urbanization has a stronger effect on urban–rural temperature differences than DPD; therefore, the dampening of urban temperature ranges by urbanization could be driving the differentiation between urban and rural station classification.
Alternatively it could be viewed that urban sites have a larger diurnal range of DPD values for given values of DTR. This finding is also consistent with established theory (Oke 1987) that shows urban sites are characteristically drier by day and moister at night. Such conditions elevate the DPD range at urban sites. Despite individual temperature indices showing greater power as discriminant predictors, suggesting that the temperature-related urban effects are more likely to be the dominant variable, the interplay between the diurnal temperature and moisture is critical. The ephemeral characteristics of urban–rural moisture differences are captured by dewpoint depression indices and are critical to creating a classification method that is applicable over a range of climates across the United States. For instance, if only DTR was used, arid urban locations such as Phoenix would appear as rural relative to other more-humid U.S. stations. The introduction of DPD as a second discriminant variable accounts for this large-scale climatic difference.
Three rural stations with unique topography or moisture sources are misclassified (Fig. 3) as urban by the discriminant classification. Binghamton, New York, (BGM) is misclassified as urban, which could be related to the hilltop location of the airport. Jin et al. (2008) found hilltops show little diurnal variability in relative humidity in comparison with foot-gulch locations while temperatures showed a little more diurnal variability in foot gulches. A comparison with nearby stations also showed a depressed DTR due to lower maximum temperatures stemming from changes in daytime airflow from the hilltop. Binghamton also showed little variability in discriminant scores over its 57-yr record, indicating the current climatic conditions have not changed. The similarly misclassified station at Lander, Wyoming, is located in an intermontane basin. The influence of diurnal variability in mountain and valley winds reflects an urban signature in temperatures and DPDs. Clear and calm nights can create nocturnal inversions and, in combination with moisture from the nearby river, cause valley fog and inhibit dewfall that maintains drier rural air in comparison with urban air.
Of the rural stations that are misclassified, Cape Hatteras, North Carolina, is perhaps the most unique climatically because of its location on a narrow barrier island with daily patterns in land and sea breezes, particularly under clear and calm conditions. These breezes are likely to influence temperatures and the moisture content of the air and contribute to a smaller range in temperatures than would be expected for a rural station, given the small range in DPD caused by its maritime location. Cape Hatteras was the only rural coastal station included in this part of the analysis, and it was suspected that coastal climatological characteristics affected the relationship between the discriminant variables.
A comparison of hourly diurnal temperatures at Cape Hatteras with an urban coastal (Norfolk, Virginia) and rural inland (Lynchburg, Virginia) station revealed the effects of urbanization and maritime influence on DTR (Fig. 4). Diurnal temperatures at coastal rural locations show less variability than those at coastal cities, which is the opposite of the expected pattern at inland sites. The high heat capacity of the adjacent ocean dampens the DTR. Because the coastal rural stations are misclassified as urban using the discriminant factors that perform well for continental stations, a separate discriminant function was developed for coastal stations.
Table 4 shows an additional 21 urban and rural stations that are located within 15 km of the coast, where temperature range is most affected (Sakakibara and Owa 2005). For these coastal stations, T1W and T6n correctly classified all urban and rural stations (Fig. 5), where the chance of these variables being selected at random was 0.5%. Urban and rural coastal stations are classified using the equation
where stations with m̂ < 2.551 are classified as urban and those with m̂ > 2.551 are classified as rural. Urban coastal sites have modified surfaces and airflows that enhance diurnal temperature ranges, as compared with rural coastal locations’ marine influence. Larger 6-h warming rates at urban stations are consistent with their enhanced diurnal temperature cycles relative to rural stations.
b. Temporal effects
When the original discriminant functions based on TD9 and DPDN were applied to data from an earlier (1960–69) time interval, inconsistent urban and rural station classification resulted. Almost 90% of stations are classified as rural in this period (Fig. 6) despite more of the stations being urban at this time. It was suspected that data inhomogeneities might have limited the application of the urban/rural classification functions during earlier periods across the set of 41 stations. To explore this possibility more closely, TD9 and DPDN indices were used to examine the temporal change in urban/rural classification for three selected stations by computing discriminant scores based on TD9 and DPDN averages from running 10-yr periods spanning each station’s period of record.
Figure 7 shows the change in discriminant scores for PHX, BWI, and LBF airports, with site and instrument changes also noted. Station location information and satellite images for Phoenix, Baltimore, and North Platte can be accessed through the NOAA/NESDIS/NCDC Multi-Network Metadata System. Although Phoenix and Baltimore both meet the criteria to be classed as urban stations, PHX airport is located in a large urban area close to the city center while BWI is situated in a lightly wooded suburban area. North Platte is a rural station. The two urban stations (Phoenix and Baltimore) generally show decreases in discriminant scores that indicate that the sites became more urban during the analysis period. These trends toward more urban environments are supported by Brazel et al. (2000). While discriminant scores at both stations decrease, and cross the threshold separating urban from rural discriminant scores, the magnitude of change at Baltimore is much smaller and the discriminant scores remain close to the urban/rural threshold. The largest decreases in the discriminant scores occur during periods of rapid urbanization, which fully encompassed the areas surrounding the airports at Phoenix and Baltimore (Brazel et al. 2000). North Platte is classified as strongly rural and shows no trend toward urbanization.
Instrument and site changes appear to affect discriminant scores at all three sites, with a number of fluctuations coinciding with documented events. Of particular note is the upward shift in discriminant scores that occurs at all three stations in association with the installation of the HO60 dial hygrometer during the early 1960s, shown in Fig. 7. In a similar way, the subsequent instrument change to the HO83 chilled-mirror hygrometer causes a subtle shift at Baltimore and North Platte but not at Phoenix. The change to the HO60 and, to a lesser degree, HO83 hygrothermometers is such that the urban locations appear more rural, consistent with the predominant shift in station classification in Fig. 6. Instrument changes that are based on different technology to calculate or measure dewpoints have an effect on the discriminant scores that is similar to that of environmental change. Without adjustment, the effect of these instrument changes on temperatures and DPD may mask the signature urban–rural differences inherent to homogenous series. Similarly, specific nonstandard instrument siting, encroachment of buildings and tarmac around the observation location, and multiple station moves can distort the classification of broader-scale urban or rural character of a specific site. Therefore, the discriminant relationships in Eqs. (1) and (2) are not valid if the instruments have changed; however, relationships specific to each type of instrument could be developed.
A preliminary investigation of the effects of the shifts detected in temperatures and dewpoints on urban/rural classification was performed for Baltimore, because this station shows a shift from rural to urban classification and contains only four documented metadata changepoints. No breakpoints were identified in the temperature series for Baltimore, but dewpoint inhomogeneities occur in 1960 and 1985 (Fig. 7). Detected breakpoints at this site could also be verified using the nearby Camp Springs Andrews Air Force Base, located 43 km from BWI, where station metadata show no coincident changes. Comparison of data from Andrews Air Force Base with BWI indicates that the effect of installing the HO83 in the mid-1980s was a substantial increase of dewpoints at Baltimore. In contrast, the introduction of the HO60 instrument showed a bias toward lower dewpoints.
As part of this preliminary investigation of inhomogeneities, discriminant scores for Baltimore were adjusted. Adjustments were calculated based on the shift in 10-yr-averaged discriminant scores. The adjusted discriminant scores for Baltimore are shown in Fig. 8. Adjusted discriminant scores cross the dividing point indicating urbanism slightly earlier for adjusted data.
Despite instrumental effects on dewpoints, the shift in discriminant scores for Baltimore indicates that urbanization occurred during the period of rapid urban expansion in the 1970s and early 1980s indicated by Brazel et al. (2000). However, the current weather station at BWI is located in a wooded area that is part of a green belt in the suburban periphery on the city. This apparently dampens the effects of urbanization. The PHX weather station is situated in a concentrated urban environment that its recent discriminant scores indicate is highly urbanized. Even without adjustment for inhomogeneities that make Phoenix look more rural, the station displays a marked change toward an urban classification in the 1970s (Fig. 7), which is coincident with a period of urban growth surrounding the climate station as shown in Brazel et al. (2000). The subtle increase in rurality indicated at North Platte during the 1960s and 1970s could also be attributed to instrument changes (Fig. 7). Drought conditions may also increase irrigation on adjacent farmland, which would dampen diurnal changes in DPD and make the station appear more rural.
In the longer records, the early data used in the discriminant analysis depicted in Fig. 6 likely violate assumptions of homogeneity. Systematic instrument and site changes distort discriminant scores toward values characteristic of present-day rural locations. Despite the data homogeneity issues present in most temperature and dewpoint datasets, the effects of urbanization tend to be presented as trends in the discriminant scores rather than sudden shifts. Identifying periods for which discriminant scores show trends toward an urban classification will aid in identifying periods during which urban influences affect data, despite shorter-term shifts toward seemingly more rural conditions that are attributable to inhomogeneities. Presumably this trend-based approach will be useful for older data for which metadata and comparable stations do not exist, especially for historical dewpoints for which ancillary evidence of urbanization such as satellite and/or aerial images is lacking.
Further investigation of the application of data homogenization techniques such as that of Brown and DeGaetano (2009) is required before this urban/rural classification technique can be used operationally because of the limited sample of three stations investigated here. Unless homogeneous data are used, the discriminant relationships developed using post-ASOS data are invalid. However, if the data or discriminant scores from the pre-ASOS periods can be adjusted, urban/rural classification based on the discriminant score thresholds given in Eqs. (1) and (2) appears possible (e.g., Fig. 8). Otherwise, the character of the trend in discriminant scores should be assessed, in light of known discontinuities (particularly the introduction of the HO60 hygrothermometer) that appear to introduce a systematic bias into the classification, and a more subjective determination of urban influences made.
The results of this study suggest that the relationship between DTR indices of temperature and dewpoint depression can be used as an indicator of urbanization. Annual averages of diurnal temperature range (TD9) and nighttime dewpoint depression range (DPDN) indices can be used to differentiate between urban and rural stations. Urban climates have a lower diurnal temperature range and higher range of nighttime dewpoint depression than do rural climates. These relationships can be also extended to earlier data records for which previously used methods for urban/rural classification either are not possible or can only be applied to discrete snapshots of time. However, during these earlier periods, changes in instrumentation and observing practice introduce inhomogeneities with magnitudes similar to the urban–rural differences that drive the classification technique.
When station metadata are available, the application of established adjustment techniques shows promise for homogenizing the indices upon which classification is based. The adjusted values then appear to be compatible with the discriminant functions derived from the recent data record. If homogenization is not possible, the magnitude of the observed trends in discriminant score provides a subjective measure of urbanization in earlier periods with consistent observation practices. Thus the method described is still valid, albeit the urban/rural threshold developed for the recent record may not be applicable.
In addition to these caveats, the potential exists that unique microclimatic features may influence the classification at some stations (e.g., BGM). Extensive areas of woodland and nearby creeks/rivers may affect the urban moisture characteristics. Except for coastal areas, for which a secondary classification exists using 1-h warming rates (T1W) and minimum forward 6-h temperature range (T6n), a systematic microclimatic bias is not present. The limited number of remaining misclassifications are such that current rural stations appear urban, an error that can be detected based on the current features surrounding a site, because transitions from urban to rural are unlikely.
The method described provides a data-driven approach to assessing urban character and identifying periods of active urbanization. These methods complement traditional methods of urban classification that tend to be static and rely on the subjective evaluation of ancillary data sources that are often unavailable in the presatellite era.
This project was supported by NOAA Grant NA07OAR4310061. Partial support was also obtained through NOAA contract EA133E07CN0090.
Corresponding author address: Paula Brown, Cornell University, 1121 Bradfield Hall, Ithaca, NY 14853. Email: firstname.lastname@example.org