THE U.S. HISTORICAL CLIMATOLOGY NETWORK MONTHLY

New bias adjustments reduce uncertainty in temperature trends for the United States.

ince 1987, the National Oceanic and Atmospheric

Administration’s (NOAA’s) National Climatic Data

Center (NCDC) has used observations from the U.S.

Historical Climatology Network (HCN) to quantify

national- and regional-scale temperature changes in the

conterminous United States (CONUS). To that end, U.S.

HCN temperature records have been “corrected” to

account for various historical changes in station loca-

tion, instrumentation, and observing practice. The HCN

is actually a designated subset of the NOAA Cooperative

Observer Program (COOP) Network—the HCN sites

having been selected according to their spatial coverage,

record length, data completeness, and historical stability.

The U.S. HCN, therefore, consists primarily of long-term

COOP stations whose temperature records have been

adjusted for systematic, nonclimatic changes that bias

temperature trends.

THE U.S. HISTORICAL CLIMATOLOGY

NETWORK MONTHLY TEMPERATURE

DATA, VERSION 2

b y Ma t t h e w J. Me n n e , Cl a u d e n. wi l l i a M s Jr., a n d ru s s e l l s. Vo s e

In support of its climate monitoring and assessment

activities, NCDC has recently developed an improved

U.S. HCN dataset (hereafter called HCN version 2). In

this paper we describe the HCN version 2 temperature

data in detail, focusing on the quality-assured dataset

sources as well as the bias adjustment techniques em-

ployed in version 2 to further reduce uncertainty in the

U.S. instrumental temperature record. The HCN bias

adjustments are discussed in the context of their effect on

U.S. temperature trends and in terms of the differences

between version 2 and its widely used predecessor (now

termed HCN version 1).

DATA. Network development. The U.S. HCN is a refer-

ence station network (Collins et al. 1999), that is, a subset

of long-term climate stations managed as part of a larger

network—in this case the COOP Network shown in Fig. 1.

Fi g . 1. Distribution of COOP

stations in the CONUS (black

dots) and the U.S. HCN version

2 sites (red triangles).

993

JULY 2009AMERICAN METEOROLOGICAL SOCIETY

The original HCN stations were iden-

tified in the mid-1980s by examining

station records (and metadata) from

the COOP Network with the goal

of maximizing record length, data

completeness, and stability in sta-

tion location (Quinlan et al. 1987).

To be designated as part of the HCN,

a COOP station was ideally required

to be active circa 1987 and to have a

period of record of at least 80 years.

In practice, these criteria were some-

times relaxed to provide a more uni-

form distribution of stations across

the country and to incorporate the

recommendations of the nation’s state

climatologists. The resulting network

contained 1,219 COOP stations, 84

of which were composites formed

using consecutive records from two or more stations

to achieve the minimum period of record goal.

The actual subset of stations constituting the HCN

has changed twice since 1987. By the mid-1990s, sta-

tion closures and relocations had already forced a

reevaluation of the composition of the U.S. HCN as

well as the creation of additional composite stations.

The reevaluation led to 52 station deletions and 54

additions, for a total of 1,221 stations (156 of which

were composites). Since the 1996 release (Easterling

et al. 1996), numerous station closures and relocations

have again necessitated a revision of the network.

As a result, HCN version 2 contains 1,218 stations,

208 of which are composites; relative to the 1996

release, there have been 62 station deletions and 59

additions.

Figure 1 depicts the locations of the 1,218 stations

in HCN version 2. Consistent with previous releases,

the spatial distribution is reasonably uniform across

the CONUS, although station density is higher across

the eastern CONUS than in the intermountain west.

Moreover, as depicted by Fig. 2, the composition of the

network is not uniform in time. For example, there is

a rapid increase in the number of stations reporting

until about 1925, with spatial coverage increasing

most prominently in the west during these early years.

The number of stations reporting remained relatively

consistent until the end of the twentieth century, after

which it has declined because of station closures.

Source data. To maximize data completeness, HCN

version 2 was derived from the following five comple-

mentary source datasets archived at NCDC:

• DSI-3200:U.S.CooperativeSummaryoftheDay,

• DSI-3206:U.S.CooperativeSummaryoftheDay

(pre-1948),

• DSI-3210:U.S.SummaryoftheDayFirstOrder

Data,

• DSI-3220:U.S.SummaryoftheMonth,and

• U.S.HCNversion1monthlydata.

The first three datasets contain daily records,

while the last two consist of monthly means. Each

source contains “estimated” values and quality

assurance (QA) flags; however, to standardize QA

across data sources, neither the estimated values nor

the quality flags were employed in building HCN

version 2. Instead, each daily data source was sub-

jected to the suite of QA reviews listed in Table 1. The

QA checks were performed in the order in which they

appear in the table, with each procedure operating

on only those values that did not fail any of the pre-

ceding tests. The thresholds were selected and the

performance of each check was evaluated using the

AFFILIATIONS: Me n n e , wi l l i a M s , a n d Vo s e —NOAA/NCDC,

Asheville, North Carolina

CORRESPONDING AUTHOR: Matthew J. Menne, NOAA/

NCDC, 151 Patton Ave., Asheville, NC 28801

E-mail: matthew.menne@noaa.gov

The abstract for this article can be found in this issue, following the

table of contents.

DOI:10.1175/2008BAMS2613.1

In final form 5 December 2008

Fi g . 2. Number of U.S. HCN stations with temperature records.

994

JULY 2009

method outlined in Durre

et al. (2008). Collectively,

the daily QA system had

an estimated false-positive

rate of 8% (i.e., the per-

cent of flagged values that

appear to be valid) and a

miss rate of less than 5%

(the percent of true errors

that remain undetected).

Monthly means were then

derived from the quality-

assured daily data, with a

requirement that no more

than nine values be f lagged

or missing in any given

month.

The five sources were

subsequently merged by

COOP station number to

form a comprehensive data-

set of serial monthly tem-

perature values. Duplicate

rec ord s b e t we en d at a

sources were eliminated

based on a simple data-

set priority scheme (i.e.,

DSI-3200hadthehigh-

est ranking, followed by

DSI-3206,andsoon).The

resulting merged dataset

was then subjected to the

three additional monthly

QA checks listed in Table 2;

together, these checks had

a false-positive rate of 15%

for maximum temperature

and 10% for minimum tem-

perature. Note that the two

spatial checks were performed after the climatological

check; furthermore, each was applied iteratively until

no additional spatial inconsistencies were detected.

The monthly QA reviews removed fewer than 0.2%

of monthly maximum and minimum temperature

values.

SOURCES AND ASSESSMENT OF TEM-

PERATURE BIAS IN THE U.S. HCN. The

process of removing systematic changes in the bias

of a climate series is called homogenization, and the

systematic artificial shifts in a series are frequently

referred to as “inhomogeneities.” In the HCN, there

are a number of causes behind inhomogeneities,

including changes to the time of observation, station

moves, instrument changes, and changes to condi-

tions surrounding the instrument site. An assessment

of each of these causes is discussed below.

Bias caused by changes to the time of observation.

The majority of the COOP Network observers (and

also HCN) are volunteers who make observations at

times that are more convenient than local midnight.

However, the time at which daily maximum and

minimum temperatures are observed has a systematic

effect on the calculation of the monthly mean (Baker

1975; Karl et al. 1986). This “time of observation bias”

would be of little concern with regard to tempera-

Ta b l e 1. Quality assurance checks applied to daily data.

Data problem Description of check

Simultaneous zeros Identifies days on which both maximum and minimum

temperature are −17.8°C (0°F)

Duplication of data Identifies duplication of data between entire years,

different years in the same month, different months within

the same year, and maximum and minimum temperature

within the same month

Impossible value Determines whether a temperature exceeds known world

records

Streak Identifies runs of the same value on >15 consecutive days

Gap Identifies temperatures that are at least 10°C warmer or

colder than all other values for a given station and month

Climatological outlier Identifies daily temperatures that exceed the respective

15-day climatological means by at least six standard

deviations

Internal inconsistency Identifies days on which the maximum temperature is less

than the minimum temperature

Interday inconsistency Identifies daily maximum temperatures that are less than

the minimum temperatures on the preceding, current, and

following days as well as for minimum temperatures that

are greater than the maximum temperatures during the

relevant 3-day window

Lag-range inconsistency Identifies maximum temperatures that are at least

40°C warmer than the minimum temperatures on the

preceding, current, and following days as well as minimum

temperatures that are at least 40°C colder than the

maximum temperatures within the 3-day window

Temporal inconsistency Determines whether a daily temperature exceeds that on

the preceding and following days by more than 25°C

Spatial inconsistency Identifies temperatures whose anomalies differ by more

than 10°C from the anomalies at neighboring stations on

the preceding, current, and following days

“Mega” inconsistency Looks for daily maximum temperatures that are less than

the lowest minimum temperature and for daily minimum

temperatures that are greater than the highest maximum

temperature for a given station and calendar month

995

JULY 2009AMERICAN METEOROLOGICAL SOCIETY

ture trends provided that the observation time at a

given station did not change during its operational

history.AsshowninFig.3,however,therehasbeen

a widespread conversion from afternoon to morning

observation times in the HCN. Prior to the 1940s,

for example, most observers recorded near sunset in

accordance with U.S. Weather Bureau instructions.

Consequently, the U.S. climate record as a whole

contains a slight positive (warm) bias during the first

half of the century. A switch to morning observation

times has steadily occurred since that time to satisfy

operational hydrological requirements. The result has

been a broad-scale reduction in mean temperatures

that is simply caused by the conversion in the daily

reading schedule of the Cooperative Observers. In

other words, the gradual conversion to morning

observation times in the United States during the

past 50 years has artificially reduced the true tem-

perature trend in the U.S. climate record (Karl et al.

1986;Voseetal.2003;HubbardandLin2006;Pielke

et al. 2007a).

To account for this time of observation bias (TOB)

in the HCN version 2 monthly temperatures, the

adjustment method described in Karl et al. (1986)

was used. The robustness of this method, which was

also used to produce version 1, has been verified by

Voseetal.(2003).Inparticular,becausetheTOB

adjustment requires documentation of changes to the

observationschedule,Voseetal.(2003)verifiedthe

accuracy of the U.S. HCN time of observation history

using an independently generated source of metadata

(DeGaetano 2000). In addition, the predictive skill of

the Karl et al. (1986) approach to estimating the TOB

was confirmed using hourly data from 500 stations

during the period 1965–

2001 (whereas the approach

was originally developed

using data from 79 stations

during the period 1957–64).

Given these verifications,

the Karl et al. (1986) TOB

adjustment procedure was

used in HCN version 2

without modification.

To calculate the effect of

the TOB adjustments on the

HCN version 2 temperature

trends, the monthly TOB-

adjusted temperatures at

each HCN station were

converted to an anomaly

relative to the 1961–90 sta-

tion mean. Anomalies were

then interpolated to the nodes of a 0.25° × 0.25°

latitude–longitude grid using the method described

by Willmott et al. (1985). Finally, gridpoint values

were area weighted into a mean anomaly for the

CONUS for each month and year. The process was

then repeated for the unadjusted temperature data,

and a difference series was formed between the TOB-

adjusted and unadjusted data, as shown in Fig. 4.

Figure 4 indicates that removing the time of

observation bias progressively elevates the mean

U.S. temperature relative to the raw value during

the period that coincides with the gradual shift to

morning observation times in the network. The net

effect of the TOB adjustments is to increase the overall

trend in maximum temperatures by about 0.015°C

decade

−1

(±0.002) and in minimum temperatures by

about 0.022°C decade

−1

(±0.002) during the period

Ta b l e 2. Quality assurance checks applied to monthly data.

Data problem Description of check

Climatological outlier Identifies temperatures that exceed their respective

climatological means for the corresponding station and

calendar month by at least five standard deviations

Spatial inconsistency Compares z scores (relative to their respective

climatological means) to concurrent z scores at the nearest

20 neighbors located within 500 km of the target; a

temperature fails if (i) its z score differs from the regional

(target and neighbor) mean z score by at least 3.5 standard

deviations and (ii) the target’s temperature anomaly differs

by at least 2.5°C from all concurrent temperature anomalies

at the neighbors

Spatial inconsistency Identifies temperatures whose anomalies differ by more

than 4°C from concurrent anomalies at the five nearest

neighboring stations whose temperature anomalies are well

correlated with the target (correlation >0.7 for the cor-

responding calendar month)

Fi g . 3. Changes in the documented time of observation

in the U.S. HCN.

996

JULY 2009

1895–2007. This net effect is about the same as that

of the TOB adjustments in the HCN version 1 tem-

perature data (Hansen et al. 2001), which is to be

expected since the same TOB-adjustment method is

used in both versions.

Bias associated with other changes in observation

practice. In addition to changes in the time of obser-

vation, most surface weather stations also experience

changes in station location or instrumentation at vari-

ous times throughout their histories. Such modifica-

tions generally entail alterations in sensor exposure

and/or measurement bias that cause shifts in the

temperature series that are unrelated to true climate

variations. In HCN version 1, the effects of station

moves and instrument changes were addressed using

the procedure described by Karl and Williams (1987).

Because this procedure addressed changes that are

documented in the NOAA/NCDC station history

archive, the HCN version 1 homogeneity algorithm

was called the Station History Adjustment Program

(SHAP).

Unfortunately, COOP station histories are in-

complete. As a result, discontinuities may occur

with no associated record in the metadata. Since

undocumented discontinuities remain undetected by

methods like SHAP, a new homogenization algorithm

was developed for the HCN version 2 temperature

data (Menne and Williams 2009). This new algorithm

addresses both documented and undocumented dis-

continuities via a pairwise comparison of temperature

records, which avoids problems associated with the

use of reference series in undocumented change-

point detection (Menne and Williams 2005). In the

pairwise approach, comparisons are made between

numerous combinations of temperature series in a

region to identify and remove relative inhomogene-

ities (i.e., abrupt changes in one station series relative

to many others).

The pairwise approach works best when there are

many neighboring series available for comparison

with each target series. Thus, to maximize the num-

ber of potential neighbors for each HCN station, all

COOP temperature series were used as input by the

pairwise algorithm. In contrast, the SHAP used in

HCN version 1 was restricted to intercomparing only

HCN series, in large part because digital monthly

COOP temperature data (and metadata) were more

limited back in the 1980s. Since that time, digitiza-

tion efforts under the Climate Data Modernization

Program (CDMP 2001) have markedly increased the

volume of digital station data and histories avail-

able for the early years of the Cooperative Observer

Program, as shown in Fig. 5. As noted in the “Data”

section, these historical temperature values were

merged with other COOP Network data sources,

which effectively increased the density of the obser-

vations (as well as the correlation between all series

tested), thereby improving the ability of the pairwise

algorithm to detect relative inhomogeneities.

As in HCN version 1, homogeneity testing in HCN

version 2 was conducted separately on monthly-

mean maximum and minimum temperature series.

Figure 6 depicts the frequency and magnitude of

shifts detected by the pairwise algorithm for each

variable. Overall, the pairwise algorithm identified

around 6,000 statistically significant changepoints

in maximum temperature series and roughly 7,000

shifts in minimum temperature series. Since there are

approximately 120,000 station years of temperatures

in the HCN version 2 dataset, this represents an aver-

Fi g . 4. Average annual differences over the CONUS

between the TOB-adjusted data and the unadjusted

(raw) data.

Fi g . 5. Digital data availability for COOP stations

before (DSI 3200) and after (DSI 3200 + 3206) the

digitization efforts of the Climate Data Moderniza-

tion Program.

997

JULY 2009AMERICAN METEOROLOGICAL SOCIETY

age of about one significant artificial shift for every

15–20 years of station data. In terms of the adequacy

of the HCN metadata, about half of the identified

inhomogeneities are undocumented.

Most of the documented changes in the HCN are

associated with station relocations. In theory, minor

station moves or other changes to sensor exposure

would be expected to have

a more pronounced effect

on minimum tempera-

tures than on maximum

temperatures. The reason

is that minimum tempera-

tures generally occur near

sunrise when calm and

stable atmospheric bound-

ary layer conditions are

prevalent, at which time

near-surface temperature

fields are strongly coupled

to the local surface charac-

teristics (Oke 1987). On the

other hand, during daylight

hours, the boundary layer

is more commonly well

mixed and microclimate

differences between nearby

locations should be less

evident. The larger number

of shifts detected in mini-

mum temperature series

relative to maximum tem-

perature series is consistent

with this reasoning.

Whereas station changes

can cause either an artificial

rise or drop in temperature,

the distribution of shifts

identified in HCN version 2

is not necessarily symmet-

ric about zero. For example,

there are about 400 more

negative shifts than posi-

tive shifts in maximum

temperature series (Fig. 6a).

Most of this asymmetry

appears to be associated

with documented changes

in the network (Fig. 6e)

and, in particular, with

shifts caused by the transi-

tion from liquid-in-glass

(LiG)thermometerstothe

maximum–minimum temperature system (MMTS;

Fig. 6g). Quayle et al. (1991) concluded that this

transition led to an average drop in maximum tem-

peratures of about 0.4°C and to an average rise in

minimumtemperaturesof0.3°Cforsiteswithno

coincident station relocation. [These averages were

subsequently used in version 1 to adjust the records

Fi g . 6. Histograms of the magnitude of changepoints (shifts) in U.S. HCN

mean monthly maximum and minimum temperature series: (a), (b) all

changepoints; (c), (d) undocumented changepoints; (e), (f) changepoints

associated with documented station changes; (g), (h) changepoints associated

with the transition from LiG thermometers to the MMTS. A negative shift

indicates that the inhomogeneity led to a decrease in the mean level of the

temperature series relative to preceding values.

998

JULY 2009

from HCN stations that converted to the MMTS,

primarily during the mid- and late 1980s (Easterling

etal.1996).]Morerecently,HubbardandLin(2006)

estimated a somewhat larger MMTS effect on HCN

temperatures and advocated for site specific adjust-

ments in general, including those sites with no docu-

mented equipment move.

Notably, the pairwise algorithm in HCN version 2

allows for such site-specific adjustments to be calcu-

lated for all types of station changes. The subsets of

changes associated with the conversion to the MMTS

are shown in Figs. 6g and 6h. The pairwise results

indicate that only about 40% of the maximum and

minimum temperature series experienced a statisti-

cally significant shift (out of ~850 total conversions to

MMTS). As a result, the overall effect of the MMTS

instrument change at all affected sites is substantially

less than both the Quayle et al. (1991) and Hubbard

andLin(2006)estimates.However,theaverageeffect

of the statistically significant changes (−0.52°C for

maximumtemperaturesand+0.37°Cforminimum

temperatures)isclosetoHubbardandLin’s(2006)

results for sites with no coincident station move.

For HCN version 2 as a whole, the combined effect

of all adjustments for documented and undocumented

temperature changes is to increase the average U.S.

trendinmaximumtemperaturesbyabout0.031°C

decade

−1

(±0.007) over the period of record relative

to the values adjusted only for the TOB (Fig. 7). In

contrast, the effect of the pairwise homogenization

algorithm on minimum temperature trends is effec-

tively zero over the period of record. As Fig. 7 indi-

cates, the most significant effect of the adjustments

on maximum temperatures begins after 1985, which

coincides with the beginning of the changeover to

the MMTS. The trend in the difference between the

fully adjusted maximum temperature data and the

TOB-adjusted data reflects the cumulative effect of

the individual instrument changes.

Although the majority of MMTS changes occurred

during the mid- and late 1980s, about 10% of HCN

stations made the switch after 1994 (the last update

to the HCN version 1 digital metadata). In addition,

a number of sites (about 5% of the network) con-

verted to the Automated Surface Observation System

(ASOS)after1992.LiketheMMTS,ASOSmaximum

temperature measurements have been shown to be

lower relative to values from previous instruments

(e.g., Guttman and Baker 1996). Such results are in

agreement with the pairwise adjustments produced in

HCN version 2; that is, an average shift in maximum

temperatures caused by the transition to ASOS in the

HCN of about −0.44°C. The combined effect of the

transition to MMTS and ASOS appears to be largely

responsible for the continuing trend in differences

between the fully and TOB-only adjusted maximum

temperatures since 1985. On the other hand, while

the effect of ASOS on minimum temperatures in the

HCN is nearly identical to that on maximum tem-

peratures (−0.45°C), the shifts associated with ASOS

are opposite in sign to those caused by the transition

to MMTS, which leads to a network-wide partial can-

cellation effect between the two instrument changes.

Undocumented changes, which are skewed in favor

of positive shifts, further mitigate the effect of the

MMTS on minimum temperatures.

Bias associated with urbanization and nonstandard siting.

In HCN version 1, the regression-based approach of

Karl et al. (1988) was employed to account for the

effect of the urban heat island (UHI) bias on tempera-

tures in the HCN (which they found to be important

for minimum temperatures only). In contrast, no

specific urban correction is applied in HCN version 2.

The reason is that adjustments for undocumented

changepoints in HCN version 2 appear to account

for much of the changes addressed by the Karl et al.

(1988) UHI correction used in HCN version 1. In fact,

as discussed in the next section, including adjust-

ments for undocumented changepoints actually has

a greater impact on minimum temperatures than the

HCN version 1 UHI correction. Moreover, adjusting

for both documented and undocumented change-

points effectively removes most of the local, unrepre-

sentative trends at individual HCN stations that may

arise from gradual changes to the environment. The

minimum temperature time series for Reno, Nevada

(Fig. 8), illustrates this effect. Specifically, the unad-

justed data suggest that the station developed a local

Fi g . 7. Average annual differences over the CONUS

between the fully adjusted (TOB + pairwise) HCN data

and the TOB-only adjusted data.

999

JULY 2009AMERICAN METEOROLOGICAL SOCIETY

trend beginning in the 1970s, possibly as a result of

a growing urban heat island influence. In contrast,

the fully adjusted HCN version 2 data indicate that

the relative trend changes have been largely removed.

(Notably, the Reno series is also characterized by

majorstepchangesduringthe1930sand1990scaused

by station relocations. Both abrupt changes were also

removed by the HCN version 2 adjustments.) For

these reasons, the average CONUS minimum tem-

peraturetrendcalculatedfromthe30%mosturban

HCN stations (based on population metadata) are

about the same as that calculated from the remaining

more rural locations (i.e., 0.071° and 0.077°C decade

−1

respectively) during the period 1895–2007.

It is important to note, however, that although

the pairwise algorithm uses a trend identification

process to discriminate between gradual and sudden

changes, trend inhomogenieties in the HCN are not

actually removed with a trend adjustment. Rather,

the pairwise approach uses a simple difference in

means in the target minus neighbor series (before and

after a step change) to estimate the magnitude of the

shift, even when there was a relative trend between

the two series (as in the case of Reno). Ideally, trend

inhomogeneities would be removed with gradual ad-

justments and step changes with abrupt adjustments.

Unfortunately, unlike rela-

tive step changes, which

occur simultaneously in

all difference series formed

between an HCN tempera-

ture series and those of

its neighbors, a trend in-

homogeneity may begin

and end at different times

with respect to its various

neighbors. This makes it

difficult to robustly identify

the true interval of a trend

inhomogeneity (Menne

and Williams 2009).

Use of a simple differ-

ence in means test does,

however, address both grad-

ual and sudden changes,

producing what arguably

approximates the “ best

objective hypothetical

climate record available

for the corrected station”

(Pielke et al. 2007b). More

generally, accounting for

both sudden and gradual

changes is critical because spurious results may occur

if only the sudden changes are corrected (e.g., Fig. 10

in Menne and Williams 2009). The reason is that, in

some cases, gradual and sudden changes may not re-

flect station moves and the effect of urbanization but

rather some kind of microclimate peculiarity, such as

the growth and removal of a single tree. In such an

instance, correcting for the sudden change, but not for

the gradual change, would likely produce unrealistic

adjusted temperature values. Even in a case such as

the Reno observations, preserving the local trend (i.e.,

not adjusting for the gradual change) would result in

a “double counting” of the UHI signal, because the

station likely experienced urbanization effects when

it was located in the city and then again after its

relocationinthemid-1930stoanairportsite(whose

surroundings became urbanized much later).

One implication of using a difference in means

test to adjust for all changepoints is that local trends

are “aliased” onto the estimates of step changes

(DeGaetano 2006). To quantify the influence of this

aliasing effect, the pairwise approach was modified

such that only abrupt shifts were removed, thereby

creating a “nonproduction” version of HCN in

which local trends were retained (see Menne and

Williams 2009 for details). In the case of minimum

Fi g . 8. (a) Mean annual unadjusted and fully adjusted minimum temperatures

at Reno, Nevada. Error bars depict a measure of the cumulative uncertainty

(95% confidence limits) in the pairwise algorithm’s bias adjustments. The

estimated uncertainty was determined using 100 Monte Carlo simulations

in which a value within the range of pairwise estimates for the magnitude of

each shift was randomly selected and used to adjust the series accordingly.

(b) Difference between minimum temperatures at Reno and the mean from

its 10 nearest neighbors.

1000

JULY 2009

temperature, the resulting distribution of docu-

mented shifts became somewhat less skewed in

favor of negative changes, while the distribution of

undocumented shifts became more skewed in favor

of positive changes (relative to the results presented

in Fig. 6). The reason for these distributional changes

is that there is an apparent and sizable preference

for relative trends between HCN stations and their

neighbors to be negative. In other words, there is a

general tendency for HCN minimum temperature

trends to be smaller relative to surrounding COOP

stations. This means that the local trend aliasing

effect, on the whole, is removing more negative than

positive trend inhomogeneities at HCN stations,

despite cases like Reno. Thus, whereas there are

apparent residual trend inhomogeneities that remain

in some HCN series, they are more likely to be nega-

tive than positive and, collectively, there appears to be

little evidence of a positive bias in HCN trends caused

by the UHI or other local changes. It should be noted,

however, that if there is a regional signal that affects a

number of stations, its effect will be largely preserved

by the homogenization procedure.

A number of recent articles have also raised

concerns about the site characteristics of U.S. HCN

stations by way of photographic documentation

(e.g., Davey ahnd Pielke 2005; Pielke et al. 2007a,b).

Moreover, there is evidence that a large fraction

of HCN sites have poor ratings with respect to the

site classification criteria used by the U.S. Climate

Reference Network (A. Watts 2008 personal com-

munication; refer also to www.surfacestations.org

In at least one study (i.e., Mahmood et al. 2006),

photographic documentation and other sources of

information regarding the exposure characteristics

of COOP and HCN sites were used to link poor

siting with measurement bias. Such evidence raises

legitimate questions about the representativeness of

temperature measurements from a number of U.S.

HCN sites. However, from a climate change perspec-

tive, the primary concern is not so much the absolute

measurement bias of a particular site but rather the

changes in that bias over time, which the TOB and

pairwise adjustments effectively address (Vose et al.

2003;MenneandWilliams2009).

The goal of the HCN version 2 adjustments (and

homogenization in general) is not to ensure that

observations conform to an absolute standard but

rather to remove the effect of relative bias changes

that occur during a station’s history of observation.

In this regard, photographic documentation, though

valuable, is most valuable when it is used to document

the timing and causes of such shifts in bias through

time. Ultimately, the magnitude of relative changes

in the bias of observations, whatever the source,

cannot be inferred from the metadata. Instead, the

effect of station changes and nonstandard instrument

exposure on temperature trends must be determined

via a systematic evaluation of the observations them-

selves (Peterson 2006), generally through relative

comparisons. Such an analysis suggests that the effect

of undocumented changes appears to be at least as

significant as documented changes in the HCN and

that homogeneity testing for both types of shifts is

critical.

Bias assessment of estimates for missing monthly tem-

perature values. As in HCN version 1, HCN version

2 provides estimates for missing monthly maximum

and minimum temperatures. Estimates are generated

using an optimal interpolation technique known in-

formallyasFILNET(shortfor“fillinthenetwork”),

which makes use of the fully adjusted temperature

values at neighboring COOP stations. In essence, the

FILNETprocedureiteratestofindanoptimalsetof

neighboring series that minimizes the confidence

limits for the difference between the target series

and the average of neighboring series (optimized

separately for each calendar month). The difference

between the target and neighbor average is used as an

offset in the interpolation to account for climatologi-

cal differences between the target and neighbors. The

FILNETtechniqueisalsousedtoestimatedataina

series where changepoints occur too close together

in time (i.e., less than 24 months apart) to reliably

estimate the magnitude of shift identified by the

pairwise algorithm.

ToassesstheperformanceofFILNET,estimates

were generated for all mean monthly maximum and

minimum temperatures in the HCN and compared

with the observed values. Specifically, both the mean

difference and the mean absolute difference between

the estimated and observed values were calculated

separately for each decade in the HCN period of

record. As shown in Fig. 9, the mean difference be-

tweentheFILNETestimatesandtheobservedvalues

is less than 0.1°C in all decades. In addition, the mean

absolutedifferencebetweentheFILNETestimates

and the observed values decreases with time as the

density of stations in the COOP Network increases.

For the period of record as a whole, the mean differ-

encebetweenFILNETestimatesandtheobserved

SiteclassificationsarebasedonamodificationofLeroy

(1999), as described in the U.S. Climate Reference Network

(2002) Site Information Handbook.

1001

JULY 2009AMERICAN METEOROLOGICAL SOCIETY

monthly values in the HCN is 0.01°C, while the mean

absolute difference is slightly less than 0.5°C. As

showninFig.10,theFILNETprocedurehasvirtually

no systematic effect on HCN temperature trends.

COMPARISON OF U.S. HCN VERSIONS

1 AND 2 MONTHLY TEMPERATURES. To

assess the basic temperature differences between

HCN versions 1 and 2 at the national scale, the annual

CONUS averages from the two datasets were com-

pared using the same gridding procedure described

in the “Sources and assessment of temperature bias

in the U.S. HCN” section. Because the HCN version

1 release provides an optional UHI correction, two

difference series were formed for each variable: (i)

HCN version 2 minus HCN version 1 (with TOB

and SHAP adjustments), and (ii) HCN version 2

minus HCN version 1 (with TOB, SHAP, and UHI

adjustments).

Figure 11 indicates that there is a decreasing trend

in the difference series for minimum temperatures

before 1970. The trend is especially evident when the

UHI adjustment is excluded from HCN version 1. The

existence of this trend can be traced to the effect the

SHAP adjustments had on minimum temperatures

in HCN version 1. Specifically, the SHAP adjustments

are limited to documented changes that have a prefer-

ence for downward shifts (Fig. 6). When these shifts

are removed, a mean warming is introduced into the

SHAP-adjusted temperature record relative to the raw

and TOB-only adjusted data (see also Hansen et al.

2001). Notably, the HCN version 1 UHI adjustment

depresses HCN temperature series as a function of

population growth, thereby indirectly compensating

for much (but not all) of the SHAP-induced warming.

In contrast, the undocumented changepoints in mini-

mum temperatures identified in HCN version 2 are

skewed in favor of positive shifts, which collectively

compensate for the negatively skewed documented

shifts (the only changes known to the SHAP). For this

reason, the HCN version 2 pairwise adjustments do

not increase the minimum temperature trend relative

to the TOB-adjusted data (Fig. 7).

Figure 11 also suggests a divergence between

HCN versions 1 and 2 temperatures after 1985, a

difference associated with the adjustments for the

MMTS instrument change in HCN version 1. As

discussed in the “Bias associated with other changes

in observation practice” section, the HCN version 1

MMTS correction appears to be too large when the

effect on the full subset of HCN sites is considered

(i.e., when stations with documented moves coinci-

dent to MMTS installation are included). However,

as Fig. 11 indicates, maximum temperatures recover

from the apparent overcorrection in version 1 after

the mid-1990s. Unfortunately, this recovery is

accidental; in fact, it appears to be a consequence of

two factors: first, the HCN version 1 metadata were

last updated with the Easterling et al. (1996) release;

second, the continued conversion to MMTS (and later

Nimbus)—as well as the introduction of ASOS—have

artificially (but unknown to SHAP) cooled maximum

temperatures to a level that currently compensates for

the HCN version 1 overcorrection.

TEMPERATURE TRENDS FROM THE U.S.

HCN. Figure 12 depicts the U.S. annual time series

for maximum, minimum, and mean [(maximum +

minimum)/2] temperature during the period 1895–

2007. In general, all variables exhibit a slight increase

Fi g . 10. Average annual differences over the CONUS

between the fully adjusted HCN data with estimates

for missing values (TOB + pairwise + FILNET) and

the fully adjusted data without missing data estimates

(TOB + pairwise).

Fi g . 9. Difference (by decade) between FILNET esti-

mates and observed monthly values at all U.S. HCN

stations.

1002

JULY 2009

untiltheearly1930s,followedbyaslightdecrease

until the early 1970s, and finally a more prominent in-

crease into the early twenty-first century. Interannual

variability is markedly lower from the mid-1950s to

the mid-1970s, the so-called benign climate period

(Bakeretal.1993).Formaximumtemperature,the

twohighestrankingyearsare1934and2006;for

minimum temperature, the two highest values

occurred in 1998 and 2006.

Table3summarizesU.S.annualandseasonal

(linear) trends in maximum and minimum tem-

perature for the raw, TOB,

and fully adjusted (TOB

+ pairwise) HCN version

2 data as well as the fully

adjusted HCN version 1

data (TOB + SHAP + UHI).

On an annual basis, the

HCN version 2 trend in

maximum temperature is

0.064°C decade

−1

, and the

trend in minimum temper-

ature is 0.075°C decade

−1

(both of which are com-

parable to the global mean

trend of ~0.060°C decade

−1

for t he sa me pe r iod).

Trends in both variables

are largest in winter and

lowest in fall, and increases

in the minimum exceed

those in the maximum in

all seasons except spring.

For reasons described in

the “Bias caused by changes

to the time of observation”

section and “Bias associated with other changes in

observation practice” section, trends in the adjusted

data always exceed those in the raw data. However,

as discussed in last section, the HCN version 2 trends

in minimum temperature are somewhat smaller than

the fully adjusted HCN version 1 trends.

InFig.13,thegeographicdistributionoflinear

trends in maximum and minimum temperatures

for the period 1895–2007 are shown both for the

adjusted HCN version 2 data and for the raw data.

Geographically,maximumtemperature(Fig.13a)

has increased in most areas except in parts of the east

central and southern regions. Minimum temperature

(Fig.13c)exhibitsthesamepatternofchange,though

the pockets of decreasing temperature are displaced

slightly to the south and west relative to maximum

temperature.Figures13band13dsuggestthattheraw

data exhibit more extreme trends as well as larger spa-

tial variability; in other words, the bias adjustments

tend to have a spatial smoothing effect on rates of

change. The reduction in the extent of negative trends

is a function of removing the time of observation bias

and of the adjustments associated with the MMTS

instrument change.

Despite the more coherent pattern, Pielke et al.

(2007a,b) argue that homogenized data are not

useful for calculating regional trends because the

homogenized series lack independence, noting, in

Fi g . 12. Time series of annual temperature anomalies from HCN version 2

averaged over the CONUS. Base period is 1961–90. The trends include 95%

confidence limits (± one standard error) that were calculated by adding the

error in the least squares regression coefficient for the series trend and a

factor quantifying the uncertainty in the adjusted temperature values (as

described in Fig. 8).

i g . 11. Average annual differences over the CONUS

between HCN version 2 and HCN version 1 (Revision

3; Easterling et al. 1996)

1003

JULY 2009AMERICAN METEOROLOGICAL SOCIETY

particular, that the site-specific

information that would have

been obtained from a well-sited,

stable station cannot be derived

retrospectively. Nonetheless,

Pielke et al. (2007b) state that the

adjusted temperature series “may

well be the best objective hypo-

thetical climate record available.”

We believe that it follows that the

adjusted series can be used to

infer patterns of climate vari-

ability and change at the surface

(which is one of the principal

motivations behind climate data

homogenization). Moreover, the

increase in interstation correla-

tion in the adjusted data relative

to the unadjusted data is negli-

Fi g . 13. Geographic distribution of linear trends in HCN version 2 temperatures for the period 1895–

2007. (a) adjusted maximum temperatures; (b) unadjusted maximum temperatures; (c) adjusted

minimum temperatures; (d) unadjusted minimum temperatures.

Ta b l e 3. U.S. annual and seasonal temperature trends (°C decade

−1

)

1895–2007 for adjusted and unadjusted temperature series.

Season Maximum temperature Minimum temperature

Fully adjusted—Version 2 (TOB + Pairwise)

Annual 0.064 0.075

Dec–Feb 0.101 0.107

Mar–May 0.082 0.066

Jun–Aug 0.044 0.067

Sep–Nov 0.025 0.054

Unadjusted (Raw)—Version 2

Annual 0.018 0.054

Adjusted for TOB only—Version 2

Annual 0.033 0.076

Fully adjusted—Version 1 (TOB + SHAP + UHI)

Annual 0.063 0.090

1004

JULY 2009

gible (accounting for the effect of shifts). It is likely

for this reason that Vose and Menne (2004) found

that the same basic relationship exists between sta-

tion density and the error in calculating the mean

U.S. temperature trend, whether unadjusted or

adjusted data are used. In addition, the Vose and

Menne (2004) assessment of the network density

required to capture the overall U.S. trend is about

an order of magnitude less than the current configu-

ration of the HCN. This suggests that temperature

observations from the HCN should be sufficient to

calculate regional trends in most areas. In any case,

all COOP temperature series are homogenized by

the HCN version 2 pairwise algorithm, which ex-

pands the pool of adjusted series beyond the HCN

subset. Consequently, if there is a concern about

the characteristics of a particular HCN site or inad-

equate station density in some areas, adjusted COOP

temperature series can supplement the HCN. This

is only one of the benefits of this unique climate

network, made possible by the efforts of dedicated

volunteers for more than a century.

SUMMARY AND CONCLUSIONS. Overall,

the collective effect of changes in observation prac-

tice at U.S. HCN stations is of the same order of

magnitude as the background climate signal (e.g.,

artificial bias in maximum temperatures is about

−0.04°C decade

−1

compared to the background

trend of about 0.06°C decade

−1

). Consequently, bias

adjustments are essential in reducing the uncer-

tainty in U.S. climate trends. The bias changes that

have had the biggest effect on the climate network

as a whole include changes to the time of observa-

tion (which affects both maximum and minimum

temperature trends) and the widespread conversion

to the MMTS (which affects primarily maximum

temperatures). Adjustments for undocumented

changes are especially important in removing bias

in minimum temperature records. Tests for undocu-

mented shifts, however, are inherently less sensitive

than in cases where the timing of changes is known

through metadata. Thus, metadata are exceedingly

valuable when it comes to adjusting and evaluating

climate trends.

Trends in the HCN version 2 adjusted series are

more spatially uniform than in unadjusted data. This

indicates that the homogenization procedures remove

changes in relative bias and that the background

climate signal is more accurately represented by

the homogenized data. It is important to point out,

however, that although homogenization generally

ensures that climate trends can be more confidently

intercompared between sites, the effect of relative

biases will still be reflected in the mean temperatures

of homogenized series. The reason is that, by conven-

tion, temperatures are adjusted to conform to the

latest (i.e., current) observing status at all stations.

This detail helps to explain why Peterson and Owen

(2005) found evidence of a systematic difference

in mean temperatures at rural versus urban HCN

stations but little evidence of a comparable differ-

ence in their homogenized trends. Moreover, while

changes in observation practice have clearly had a

systematic effect on average U.S. temperature trends,

homogeneity matters most at the station level where

even one change in bias can have a drastic effect on

the series trend (which can occasionally be missed

by changepoint tests). Therefore, the goal behind the

HCN version 2 dataset (and future improvements) is

to make the adjustments as site specific and compre-

hensive as possible, which is especially valuable in the

development of widely used products, such as the U.S.

Climate Normals.

Finally, the U.S. HCN data will be updated

monthly and fully reprocessed periodically to detect

and adjust for shifts from the recent past (see www.

ncdc.noaa.gov/oa/climate/research/uschcn/ for fur-

ther information, including access to the data and

uncertainty calculations). Plans are also in place to

ensure that U.S. HCN monthly means are internally

consistent with NCDC’s global daily dataset (the

Global Historical Climatology Network—Daily

dataset). Still, there is always room for improvement

in the field of climate data homogenization. For

example, although the monthly adjustments used

in HCN version 2 are constant for all months, there

is evidence that bias changes often have effects that

vary seasonally and/or synoptically (Trewin and

Trivitt 1996; Guttman and Baker 1996). As shown

by Della-Marta and Wanner (2006), it is possible to

estimate the differential effects indirectly by evalu-

ating the magnitude of change as a function of the

frequency distribution of daily temperatures. Daily

adjustments are thus a promising area for future

HCN development.

ACKNOWLEDGMENTS. The authors wish to thank

Anthony Watts for his considerable efforts in documenting

the current site characteristics of U.S. HCN stations. The

authors also thank Tom Peterson, Tami Houston, and three

anonymous reviewers whose helpful comments greatly im-

proved this manuscript. Partial support for this work was

provided by the Office of Biological and Environmental

Research, U.S. Department of Energy (Grant DE-AI02-

96ER62276).

1005

JULY 2009AMERICAN METEOROLOGICAL SOCIETY

REFERENCES

Baker, D. G., 1975: Effect of observation time on

mean temperature estimation. J. Appl. Meteor., 14,

471–476.

—

,D.L.Ruschy,andR.H.Skaggs,1993:Agriculture

and the recent benign climate. Bull. Amer. Meteor.

Soc., 74, 1035–1040.

CDMP, 2001: Annual report. Climate Database Mod-

ernization Program Rep., 8 pp. [Available online at

www.ncdc.noaa.gov/oa/climate/cdmp/files/annual-

report2001.pdf.]

Climate Reference Network, 20 02: Site information hand-

book.NOAA/NESDISCRNSeriesX030,CRNRep.

NOAA-CRN/OSD-2002-0002R0UD0. [Available

online at ftp://ftp.ncdc.noaa.gov/pub/data/uscrn/

documentation/program/X030FullDocumentD0.

pdf.]

Collins, D. A., S. Johnson, N. Plummer, A. K. Brewster,

and Y. Kuleshov, 1999: Re-visiting Tasmania’s refer-

ence climate stations with a semi-objective network

selection scheme. Aust. Meteor. Mag., 48, 111–122.

Davey, C. A., and R. A. Pielke Sr., 2005: Microclimate

exposure of surface-based weather stations. Bull.

Amer. Meteor. Soc., 86, 497–504.

DeGaetano, A. T., 2000: A serially complete simu-

lated observation time metadata file for U.S. daily

Historical Climatology Network stations. Bull. Amer.

Meteor. Soc., 81, 49–67.

—

, 2006: Attributes of several methods for detecting

discontinuities in mean temperature series. J.

Climate, 19, 838–853.

Della-Marta, P. M., and H. Wanner, 2006: A method of

homogenizing the extremes and mean of daily tem-

perature measurements. J. Climate, 19, 4179–4197.

Durre, I., M. J. Menne, and R. S. Vose, 2008: Strategies

for evaluating quality assurance procedures. J. Appl.

Meteor. Climatol., 47, 1785–1791.

Easterling, D. R., T. R. Karl, E. H. Mason, P. Y. Hughes,

and D. P. Bowman, cited 1996: United States Histori-

cal Climatology Network (U.S. HCN) monthly tem-

peratureandprecipitationdata.ORNL/CDIAC-87,

NDP-019/R3.[Availableonlineathttp://cdiac.ornl.

gov/epubs/ndp019/ndp019.html.]

Guttman, N. B., and C. B. Baker, 1996: Exploratory

analysis of the difference between temperature

observations recorded by ASOS and conventional

methods. Bull. Amer. Meteor. Soc., 77, 2865–2873.

Hansen,J.,R.Ruedy,M.Sato,M.Imhoff,W.Lawrence,

D. Easterling, T. Peterson, and T. Karl, 2001: A closer

look at United States and global surface temperature

change. J. Geophys. Res., 106, 23947–23963.

Hubbard,K.G.,andX.Lin,2006:Reexaminationof

instrument change effects in the U.S. Historical

Climatology Network. Geophys. Res. Lett., 33,

L15710,doi:10.1029/2006GL027069.

Karl, T. R., and C. N. Williams Jr., 1987: An approach to

adjusting climatological time series for discontinuous

inhomogeneities. J. Climate Appl. Meteor., 26,

1744–1763.

—

, C. N. Williams Jr., P. J. Young, and W. M. Wendland,

1986: A model to estimate the time of observation bias

associated with monthly mean maximum, mini-

mum, and mean temperature for the United States.

J. Climate Appl. Meteor., 25, 145–160.

—

, H. F. Diaz, and G. Kukla, 1988: Urbanization:

Its detection and effect in the United States climate

record. J. Climate, 1, 1099–1123.

Leroy,M.,1999:Classificationd’unsite.Météo-France,

Direction des Systèmes d’Observation. Tech. Note

35,12pp.

Mahmood,R.,S.A.Foster,andD.Logan,2006:The

geoprofile metadata, exposure of instruments, and

measurement bias in climatic record revisited. Int. J.

Climatol., 26, 1091–1124.

Menne, M. J., and C. N. Williams Jr., 2005: Detection

of undocumented changepoints using multiple test

statistics and composite reference series. J. Climate,

18, 4271–4286.

—

, and

—

, 2009: Homogenization of temperature

series via pairwise comparisons. J. Climate, 22,

1700–1717.

Oke, T. R., 1987: Boundary Layer Climates. 2nd ed.

Routledge,435pp.

Peterson, T. C., 2006: Examination of potential biases

in air temperature caused by poor station locations.

Bull. Amer. Meteor. Soc., 87, 1073–1080.

—

, and T. W. Owen, 2005: Urban heat island

assessment: Metadata are important. J. Climate, 18,

2637–2646.

Pielke, R. A., Sr., and Coauthors, 2007a: Documentation

of uncertainties and biases associated with surface

temperature measurement sites for climate change

assessment. Bull. Amer. Meteor. Soc., 88, 913–928.

—

, and Coauthors, 2007b: Unresolved issues with

the assessment of multidecadal global land surface

temperature trends. J. Geophys. Res., 112, D24S08,

doi:10.1029/2006JD008229.

Quayle, R. G., D. R. Easterling, T. R. Karl, and P. Y.

Hughes, 1991: Effects of recent thermometer changes

in the Cooperative Station Network. Bull. Amer.

Meteor. Soc., 72, 1718–1723.

Quinlan, F. T., T. R. Karl, and C. N. Williams Jr., 1987:

United States Historical Climatology Network

1006

JULY 2009

(HCN) serial temperature and precipitation data.

NDP-019, Carbon Dioxide Information Analysis

Center,OakRidgeNationalLaboratory,U.S.Depart-

ment of Energy, Oak Ridge, TN.

Trewin, B. C., and A. C. F. Trevitt, 1996: The devel-

opment of composite temperature records. Int. J.

Climatol., 16, 1227–1242.

Vose, R. S., and M. J. Menne, 2004: A method to de-

termine station density requirements for climate

observing networks. J. Climate, 17, 2961–2971.

—

, C. N. Williams Jr., T. C. Peterson, T. R. Karl, and

D.R.Easterling,2003:Anevaluationofthetimeof

observation bias adjustment in the U.S. Historical

Climatology Network. Geophys. Res. Lett., 30, 2046,

doi:10.1029/2003GL018111.

Willmott, C. J., C. M. Rowe, and W. D. Philpot, 1985:

Small-scale climate maps: A sensitivity analysis of

some common assumptions associated with grid-

point interpolation and contouring. Amer. Cartogr.,

12, 5–16.

1007

JULY 2009AMERICAN METEOROLOGICAL SOCIETY