Climatic and oceanographic factors affecting daily patterns of juvenile sea turtle cold-stunning in Cape Cod Bay, Massachusetts

Abstract

We examined the climatic factors that may affect the temporal patterns of juvenile sea turtle cold-stunning and whether local extent and temporal scale oceanographic and climatic factors that induce cold-stunning are different for different species. Using classification tree models, we demonstrate that juvenile Kemp's ridley (Lepidochelys kempii) and loggerhead (Caretta caretta) sea turtles cold-stun under slightly differing oceanographic and climatic conditions within any given year. In Cape Cod Bay, Massachusetts, cold-stunned juvenile Kemp's ridley sea turtles are recovered with greater frequency (55%) during November, while the vast majority of juvenile cold-stunned loggerhead sea turtles (79%) are recovered in December. Our classification tree models suggest cold-stunned juvenile Kemp's ridleys are more often recovered from 9 November to 9 December on days with sea surface temperatures between 7.0 and 10.4 degrees C, wind speeds exceeding 5.3 m/s, air temperatures below 10.4 degrees C, and barometric pressures exceeding 1009.5 mm. Our models also suggest cold-stunned juvenile loggerheads are recovered after 5 December on days with sea surface temperatures between 5.6 and 9.0 degrees C, wind speeds exceeding 7.6 m/s, and barometric pressure exceeding 1015.9 mm. Mean straight carapace lengths (SCL) differed for the two species, Kemp's SCL=26.9 cm (n=218, range 18.4-37.2), and loggerheads SCL=52.5 cm (n=54, range 40.0-89.6). As a result, the larger sized loggerheads were able to withstand colder sea surface temperatures for longer periods of time due to greater thermoregulatory capabilities. These results demonstrate the seasonality of juvenile sea turtle cold-stunning in Cape Cod Bay, Massachusetts, providing oceanographic and climatic thresholds for the Sea Turtle Rescue and Salvage Network to maximize recovery efforts during peak cold-stunning conditions.

Full text

870 CHELONIAN CONSERVATION AND BIOLOGY, Volume 4, Number 4 – 2005
Chelonian Conservation and Biology, 2005, 4(4):870–877
© 2005 by Chelonian Research Foundation
Climatic and Oceanographic Factors Affecting Daily Patterns of
Juvenile Sea Turtle Cold-Stunning in Cape Cod Bay, Massachusetts
BRETT M. STILL1, CURTICE R. GRIFFIN1, AND ROBERT PRESCOTT2
1Department of Natural Resources Conservation, University of Massachusetts,
Amherst, Massachusetts 01003 USA [E-mail: brett_still@hotmail.com];
2Massachusetts Audubon Society, Wellfleet Bay Wildlife Sanctuary,
P.O. Box 263, South Wellfleet, Massachusetts 02663 USA
ABSTRACT. – We examined the climatic factors that may affect the temporal patterns of juvenile sea
turtle cold-stunning and whether local extent and temporal scale oceanographic and climatic factors
that induce cold-stunning are different for different species. Using classification tree models, we
demonstrate that juvenile Kemp’s ridley (Lepidochelys kempii) and loggerhead (Caretta caretta) sea
turtles cold-stun under slightly differing oceanographic and climatic conditions within any given
year. In Cape Cod Bay, Massachusetts, cold-stunned juvenile Kemp’s ridley sea turtles are
recovered with greater frequency (55%) during November, while the vast majority of juvenile cold-
stunned loggerhead sea turtles (79%) are recovered in December. Our classification tree models
suggest cold-stunned juvenile Kemp’s ridleys are more often recovered from 9 November to 9
December on days with sea surface temperatures between 7.0 and 10.4ºC, wind speeds exceeding 5.3
m/s, air temperatures below 10.4ºC, and barometric pressures exceeding 1009.5 mm. Our models
also suggest cold-stunned juvenile loggerheads are recovered after 5 December on days with sea
surface temperatures between 5.6 and 9.0ºC, wind speeds exceeding 7.6 m/s, and barometric
pressure exceeding 1015.9 mm. Mean straight carapace lengths (SCL) differed for the two species,
Kemp’s SCL = 26.9 cm (n = 218, range 18.4–37.2), and loggerheads SCL = 52.5 cm (n = 54, range 40.0–
89.6). As a result, the larger sized loggerheads were able to withstand colder sea surface temperatures
for longer periods of time due to greater thermoregulatory capabilities. These results demonstrate
the seasonality of juvenile sea turtle cold-stunning in Cape Cod Bay, Massachusetts, providing
oceanographic and climatic thresholds for the Sea Turtle Rescue and Salvage Network to maximize
recovery efforts during peak cold-stunning conditions.
KEY WORDS. – Reptilia; Testudines; Cheloniidae; Lepidochelys kempii; Caretta caretta; sea turtle;
CART; classification tree modeling; cold-stunning; Cape Cod Bay; Massachusetts; USA
Portions of coastal New England waters are considered
the northern-most developmental habitats for juvenile
Kemp’s ridley (Lepidochelys kempii) and loggerhead (Caretta
caretta) sea turtles along the U.S. Atlantic coast (Bleakney,
1965; Lazell, 1976, 1980; Morreale et al., 1992; Shoop and
Kenney, 1992; Morreale and Standora (In press). Morreale
and Standora (In press) hypothesized that these food-rich
developmental habitats are crucial for Kemp’s ridley and
loggerhead sea turtles as they shift from pelagic to early
juvenile stages. In Long Island Sound, New York, Morreale
and Standora (1994) reported measurable, and in some cases
substantial, growth in carapace length for recaptured juve-
nile sea turtles over a three-month period from July through
September. Yet, these northern developmental habitats can
only be utilized for a relatively short period of time each year
due to limited thermoregulatory capabilities of juvenile
cheloniid sea turtles (Spotila et al., 1997).
During the fall, sea turtle species that utilize summer
developmental habitats in northern temperate waters, such
as New England, Long Island Sound, and Chesapeake Bay,
migrate south along the coast with the onset of declining
water temperatures (Henwood and Ogren, 1987; Keinath,
1993; Musick et al., 1994; Epperly et al., 1995). If these
turtles do not migrate south early enough, they can be cold-
stunned by rapidly dropping water temperatures in late fall
and winter months.
Further, periodic cold fronts that produce rapidly drop-
ping air temperatures and strong northeast winds are a
common feature of the transition from summer to fall in New
England. Along with the cold air temperatures and high
winds, these storms trigger dropping sea surface tempera-
tures, especially in shallow coastal embayments that are
inhabited by juvenile sea turtles. These climatic and
oceanographic conditions, in combination with the semi-
enclosed embayment of Cape Cod, contribute to the
annual occurrence of juvenile sea turtle cold-stunning
events along the shores of Cape Cod Bay during Novem-
ber and December.
Previous studies of juvenile Kemp’s ridley and logger-
head cold-stunning focused primarily on water temperature
and prevailing wind direction as factors affecting the timing
and recovery location of cold-stunned turtles. Schwartz
(1978) reported juvenile turtles exhibit floating and lethar-
gic behaviors at water temperatures between 9 and 13ºC,
with death occurring at temperatures ranging from 5.0 to
6.5ºC. Witherington and Ehrhart (1989) and Burke et al.
(1991) reported similar cold-stunning water temperature
values, while also reporting that the prevailing wind direc-

871STILL ET AL. — Climatic Factors and Cold-Stunned Ridleys in Cape Cod Bay
tion was a dominant factor determining the recovery loca-
tion of cold-stunned turtles.
While these studies do mention cold-front storms, or
periods of unusually cold weather associated with cold-
stunning events, few examined a variety of climatic and
oceanographic factors. Thus, the objective of this study was
to evaluate the local scale climatic and oceanographic fac-
tors that affect the date and location of juvenile sea turtle
cold-stunning events in Cape Cod Bay, Massachusetts. We
hope this information will facilitate more effective recovery
efforts for cold-stunned sea turtles.
METHODS
Study Area. — Cape Cod Bay, the southern terminus of
the Gulf of Maine, is a 1100 km2 semi-enclosed embayment
averaging 30 m in depth, with the deepest part located near
the mouth of the bay off Provincetown (55 m). Bottom
substrates vary throughout the bay; sand bottom predomi-
nates but rock and eelgrass (Zostera marina) also commonly
occur. Many shallow tidal and inter-tidal sand flats, extend-
ing up to 2.4 km offshore in a very gentle slope, are found
from Dennis on the mid-Cape and extending eastward and
northward Truro shoreline. There are two prominent rocky
areas within the bay, one between Sesuit Harbor and Corpo-
ration Beach in Dennis, extending seaward approximately 3
km; a second more extensive area along the western shore
from Manomet to Ellisville. Extensive eelgrass beds occur
below mean low water off the Wellfleet and Truro shores.
The most extensive eelgrass beds are found along Billingsgate
Shoals and extend southwest of Wellfleet Harbor’s entrance
toward Sesuit Harbor in Dennis.
Cold-Stunned Turtle Data. — From the efforts of the
Sea Turtle Stranding and Salvage Network (STSSN), data
were available for cold-stunned turtles found on Cape Cod
Bay beaches from 1979 through 2001. These data included
information on species, date, location of stranding, and size
(straight carapace length; SCL) for each cold-stunned
turtle. Although all cold-stunned turtles were collected
with the assistance of numerous volunteers, the proto-
cols established by the STSSN help to ensure that all
potential beaches are surveyed. Thus, we believe that the
vast majority of turtles that cold-stunned in any given
year were recovered.
For all analyses, any turtle records with missing data
(species, date, or location) were omitted. For the classifica-
tion tree analyses, we defined the cold-stunning season as 1
November – 31 December each year. Thus, any records of
turtles recovered after 31 December were also omitted from
these analyses. In addition, the turtle stranding data for
1979–83 were not included in the classification tree analyses
because of uneven beach surveys and limited availability of
sea surface temperature and climatic buoy data for those early
years. Therefore, 914 turtles that stranded as a result of cold-
stunning from 1984–2001, were included in our analyses.
To examine both the oceanographic and climatic condi-
tions associated with cold-stunning events, we developed a
binary response variable (0 = absent, 1 = present) to repre-
sent the daily occurrence of juvenile cold-stunned turtles
washing ashore for both Kemp’s ridley and loggerhead sea
turtles, the two most common species to cold-stun in Cape
Cod Bay. We performed these analyses for both species to
determine if different cold-stunning conditions could be
detected for the two species. Juvenile cold-stunned green
(Chelonia mydas) sea turtles were not included in the models
due to their relatively low recovery rate with only 30 (2.3%)
recovered from 1979 to 2001.
Buoy Data. — We collected oceanographic and cli-
matic data from the National Oceanographic and Atmo-
spheric Administration (NOAA) buoy located 14 nautical
miles east of Boston Harbor from November through De-
cember (1984–2001) (http://www.ndbc.noaa.gov). These
data were used to represent conditions in Cape Cod Bay, due
to the absence of any long term monitoring stations located
within the bay. The hourly data recorded at the Boston buoy
were averaged to obtain daily values for surface water
temperature, air temperature, wind speed, and barometric
pressure. We then averaged these daily values to represent
the conditions over the three days prior to turtles washing
ashore. This was done to lessen the effect on the data when
turtles were recovered from beaches one or two days after
the stranding conditions had passed. Hourly prevailing wind
direction data were compiled to calculate a mean daily wind
direction vector using Oriana software (Kovach, 1994). This
mean daily vector was also averaged to represent the prevail-
ing wind direction over the previous three days.
Model Statistics. — We used classification tree models
to analyze the sea turtle cold-stunning and buoy data. These
models are particularly useful as a data-mining tool for
complex data with non-linear relationships, complex inter-
actions, and missing values that are common in ecological
data sets (De’ath and Fabricius, 2000). The main function of
classification tree analysis is to explain the variation found
in a single categorical response variable using either cat-
egorical or numeric explanatory variable(s) by developing a
graphical binary recursive partitioning based ‘tree’ (Breiman
et al., 1984; De’ath and Fabricius, 2000). The process begins
with the undivided data in the ‘root node’, and subsequently
partitioning the data into two homogeneous groups using
simple binary splitting rule(s) based on the explanatory
variable(s). Each split or ‘branch’ produces two ‘nodes’ that
attempt to minimize the misclassification rate of the
previous ‘node’. Each of these ‘nodes’ is then split
recursively until the maximum ‘tree’ is grown, where
each ‘terminal-node’ or ‘leaf’ contains one case. The tree
is then pruned back to an appropriate size to fit the data.
Several splitting criteria and pruning methods were de-
veloped and explained by Breiman et al. (1984) and
De’ath and Fabricius (2000).
For our classification tree analyses, the categorical
response or dependent variable was the presence (1) or
absence (0) of a cold-stunned Kemp’s ridley or loggerhead
each day of the cold-stunning season (1 November – 31
December) for each year from 1984 to 2001 (n = 1098 days).

872 CHELONIAN CONSERVATION AND BIOLOGY, Volume 4, Number 4 – 2005
We used 6 explanatory variables for each model, represent-
ing the oceanographic and climatic dynamics for each day of
the cold-stunning season (Table 1A).
We built the classification tree models using program
CART 4.0 (Salford Systems, 1998) with the following
criteria. We ran a series of fifty 10-fold cross validations
using gini splitting criteria, with the appropriate tree size
selected using the minimum rule (Breiman et al., 1984;
De’ath and Fabricius, 2000). We plotted the cross-validation
relative error versus the tree sizes produced for each of the
50 runs to develop a histogram of tree sizes where the modal
tree size was selected as the most appropriate tree. We
smoothed the cross-validation relative error curve by aver-
aging the relative error for each cross-validation over the 50
runs (De’ath and Fabricius, 2000). We labeled the tree
diagrams following the procedures of De’ath and Febricus
(2000). Each split (non-terminal node) was labeled with the
variable and its value that determined the split. Each ‘leaf’ or
‘terminal node’ was labeled with the ‘terminal node’ number
(in parentheses), the dominant classification either present (1)
or absent (0), the percentage of observations in the dominant
class, and the number of observations, respectively.
RESULTS
Cold-Stunning. — In Cape Cod Bay, Massachusetts, a
total of 1280 juvenile cold-stunned sea turtles (984 Kemp’s,
266 loggerheads, and 30 greens) were recovered from 1979
to 2002 (annual mean = 53, range 6–277 turtles/yr). The
majority of cold-stunnings occurred during November and
December, with 85% recovered from 12 November to 17
December and relatively few dead turtles recovered from
January to March the following year (Fig. 1). Juvenile cold-
stunned loggerheads were generally recovered later in the
cold-stunning season with 79% of all recoveries occurring in
December, whereas the majority (55%) of cold-stunned
Kemp’s ridleys were recovered in November. The average
SCL of Kemp’s was smaller (mean = 26.9 cm, n = 218, range
18.4–37.2 cm) than that of loggerheads (mean = 52.5 cm, n
= 54, range 40.0–89.6 cm).
The initiation of the November and December cold-
stunning season is triggered by the decline of sea surface
temperature (SST) in Cape Cod Bay, with juvenile Kemp’s
ridleys cold-stunning as the SST drops to approximately
10ºC (Fig. 2). The daily pattern of cold-stunning, however,
is associated with the timing of cold-front storms passing
through the region. During the 1999 cold-stunning season,
for example, several strong cold-front storms generated
dramatic increases in the number of cold-stunned turtles
recovered per day (Fig. 2). These cold-front storms typically
produce rapidly dropping air temperatures and increasing
wind speeds, resulting in cold-stunned turtles washing ashore
on windward facing beaches.
Prevailing Wind Direction. — The prevailing wind
direction associated with November and December cold-
front storms is the most important factor in determining the
beach recovery location of cold-stunned turtles. The overall
daily mean prevailing wind direction during November and
December from 1984 to 2002 was significantly oriented in
an easterly direction, with a mean of 94.0º (n = 1098, range
= 46–131º, circular s.d. = 70.5º, Rayleigh test of uniformity
p = 0.00). These prevailing wind data correspond to the
eastern shoreline of Cape Cod Bay where cold-stunned
turtles are recovered each year, with 90% of all cold-stunned
turtle recoveries occurring along the beaches of Truro to the
north, and Dennis to the south (Fig. 3).
Classification Tree Models
Kemp’s Ridleys. — The modal tree size for the daily
Kemp’s ridley presence absence model, selected by the
minimum rule after a series of fifty 10-fold cross validations,
contained 3-terminal nodes. This 3-terminal node classifica-
A. VARIABLE DESCRIPTION VARIABLE REL.
IMP.
P3DWTMP 1 00
DAY 62.4
P3DWTMP Avg. sea surface t emp. recorded
over the previous three -days P3DATMP 24.6
P3DPDIR 1.6 DAY Julian date of t he cold-stunning
season P3DWSPD 0.8
P3DBAR 0.3
P3DATMP Avg. air te mp. recorded hourly
over the previous three days
VARIABLE REL.
IMP.
P3DBAR
Avg. barometr ic pressure recorded
hourly over the previ ous three
days DAY 100
P3DATMP 27 P3DWSPD Avg. wind speed recorded hourly
over the previous three days (m/s) P3DPDIR 0.7
P3DWSPD 0.5
P3DPDIR
Avg. prevailing wind direction
recorded over the previous three
days
B.
C.
P3DBAR 0.0
D. VARIABLE REL.
IMP.
E. VARIABLE REL.
IMP.
F. VARIABLE REL.
IMP.
DAY 100 P3DWTMP 100 P3DWTMP 100
P3DATMP 30.8 DAY 70 DAY 70
P3DWSPD 18.3 P3DATMP 56.9 P3DATMP 61
P3DBAR 11.3 P3DPDIR 12.7 P3DWSPD 30.1
P3DPDIR 2.3 P3DWSPD 2.0 P3DBAR 10.5
P3DBAR 0.8 P3DPDIR 5.2
Table 1. Explanatory variables used in the Kemp’s ridley and
loggerhead presence/absence classification tree analyses (A).
Relative importance of the explanatory variables included in
the Kemp’s ridley and loggerhead presence/absence classifi-
cation trees (B, C, D, E, F).
0
20
40
60
80
Oct Nov Dec Jan Feb Mar Apr
Month
Percent of Cold-Stunnings
Kem p's r idle y
loggerhead
Figure 1. The highest percentages of juvenile sea turtle cold-
stunnings in Cape Cod Bay are documented during November and
December (1979–2002), with 85% of all Kemp’s and 79% of all
loggerhead cold-stunning occurring in December.

873STILL ET AL. — Climatic Factors and Cold-Stunned Ridleys in Cape Cod Bay
tion tree had an overall misclassification rate of 35.6%, and
misclassification rates of 13.9% and 43.3%, respectively for
presence (1) and absence (0) (Fig. 4A). Of the 6 explanatory
variables included in the model (Table 1A), the previous
three-day sea surface temperature (P3DWTMP) was chosen
as the primary splitting variable for the two splits in the tree.
The presence (1) of cold-stunned Kemp’s was predominant
in the second terminal node, with the tree suggesting Kemp’s
were most likely to cold-stun at sea surface temperatures
between 7.5 and 10.4ºC (Fig. 5).
The relative importance of the explanatory variables
included in the model indicated that sea surface temperature
(P3DWTMP) was the most dominant (100) of the variables
included, followed by the day of the cold-stunning season
(DAY) (62.4) and average air temperature (P3DATMP)
(24.6) having moderate importance (Table 1B). The prevail-
ing wind direction (1.6), average wind speed (0.8) and
average barometric pressure (0.3) had little relative impor-
tance in comparison to the previous three variables.
To examine the climatic factors without the influence of
oceanographic temperature dynamics, we re-ran the analy-
sis removing the dominant splitting variable (P3DWTMP).
Using the same model building criteria as the previous tree,
we obtained a modal tree size of 3-terminal nodes; however,
strong support was also shown for an 8-terminal node tree
being selected only 2% less than the modal tree size (Fig 4B).
Both splits on the 3-terminal node tree used the date of the
cold-stunning season (DAY) as the splitting variable, sug-
gesting Kemp’s are more likely to cold-stun between 9
November and 20 December (Fig. 6). Overall, the tree had
a relatively high misclassification rate of 46.1%, with
misclassification rates of 13.3 and 49.7%, respectively, for
presence (1) and absence (0) of cold-stunned Kemp’s (Fig.
6). The relative importance of the climatic variables used in
the model after removing the dominant sea surface tempera-
Figure 2. As the sea surface temperature (SST) in Cape Cod Bay drops below approx. 10ºC, juvenile cold-stunned Kemp’s ridleys begin
cold-stunning and wash up on eastern Cape Cod Bay beaches, while juvenile loggerheads typically do not cold-stun until the SST drops
to approx. 9.0ºC. The daily pattern of juvenile sea turtle cold-stunning is driven by the occurrence of cold front storms. Large spikes in the
number of cold-stunned turtles washing ashore occur with dramatic declines in air temperature (ºC) (ATMP) and increasing wind speed
(m/s) (WSPD). These daily patterns are clearly evident during the 1999 cold-stunning season.
Figure 3. Juvenile cold-stunned sea turtles are primarily recovered
along the eastern shoreline of Cape Cod Bay between Truro to the
north, and Dennis to the south (percent recovered per area). These
recovery locations correspond to the mean prevailing wind direc-
tion recorded during the cold-stunning season from 1984–2001
(mean = 94.03º).

874 CHELONIAN CONSERVATION AND BIOLOGY, Volume 4, Number 4 – 2005
ture variable (P3DWTMP) suggested the day of the cold-
stunning season (DAY) has the greatest influence (100)
followed by average air temperature (P3DATMP) with mod-
erate support (27). The remaining three climatic variables had
relatively little importance in the model (Table 1C).
The 8-terminal node tree (with P3DWTMP removed)
used four splitting variables, predicting the presence of cold-
stunned Kemp’s in three of the eight terminal nodes (termi-
nal nodes 3, 5, and 7). Further, this model yielded three
slightly different cold-stunning scenarios (Fig. 7). All three
cold-stunning scenarios began similar to the 3-terminal node
tree that indicated that Kemp’s cold-stunned between 9
November and 20 December. Additionally, scenario one
(terminal node 3) suggested cold-stunned Kemp’s were
predominantly present before 9 December (DAY) when
the average wind speed (P3DWSPD) was ≥ 5.3 m/s, and
the average air temperature (P3DATMP) was ≤ 10.4ºC
(Fig. 7). The second cold-stunning scenario (terminal
node 5) suggested that cold-stunned Kemp’s were present
under the same average wind speed and air temperature
as scenario one, and after 9 December when the baromet-
ric pressure was > 1009.5 mm (Fig. 7). The final scenario
(terminal node 7) showed Kemp’s cold-stunning with the
same wind speed (PEDWSPD) (≥ 5.3 m/s) as scenario
one and two, and with cold-stunning after 24 November
(DAY) when average air temperatures (P3DATMP) were
≥ 10.4ºC (Fig. 7). The overall misclassification rate for
this 8-terminal node tree was 38.2%, with
misclassification rates of 11.5 and 47.6%, respectively,
for presence (1) and absence (0) of cold-stunned Kemp’s.
The relative importance of the climatic variables used in
the 8-terminal node classification tree was dominated
(100) by the day of the cold-stunning season (DAY)
followed by the average air temperature (P3DATMP)
(30.8) and average wind speed (P3DWSPD) (18.3). The
average barometric pressure (P3DBAR) showed moder-
ate importance (11.3), with the average prevailing wind
direction (P3DPDIR) showing little effect (2.3) on the
model (Table 1D).
Loggerheads. — Our loggerhead presence/absence
models also had strong support for a range of tree sizes. The
modal tree size (2-terminal nodes) was selected by 22% of
the fifty 10-fold cross-validations; however, three different
tree sizes (5, 6, and 7-terminal nodes) exhibited the same
prevalence, being selected 14% out of the 50 runs. Addition-
0
0.2
0.4
0.6
0.8
1
1.2
1234578101214124
Tr ee Siz e
Corss-Validated
Relative Error
0
0.2
0.4
0.6
0.8
1
1.2
12567 81115182071
Tr ee Size
Cross-Validated
Relative Error
0
0.2
0.4
0.6
0.8
1
1.2
1234567111316114
Tree Size
Cross-Validated
Relative Error
Figure 4. 10-fold cross-validated relative error curve averaged
over 50 runs with 1 S.E. error bars for Kemp’s ridley (A and B) and
loggerhead (C) presence/absence models. The columns represent
the relative frequency of tree sizes selected by the 50 runs.
A
B
C
Figure 5. 3-terminal node classification tree model predicting the
daily presence/absence of juvenile cold-stunned Kemp’s ridley sea
turtles in Cape Cod Bay during the cold-stunning season. Each
terminal node is labeled with the node number (in parentheses), the
dominant classification (1 = presence, 0 = absence), the percent
dominance, and the number of cases in each terminal node, respec-
tively. The two splits on the tree suggest Kemp’s will cold-stun
when Cape Cod Bay sea surface temperatures (P3DWTMP) are
between 7 and 10.4ºC. Overall misclassification = 35.6%, category
0 misclassification = 43.3%, category 1 misclassification = 13.9%.
Figure 6. 3-terminal node classification tree model predicting the
daily presence/absence of juvenile cold-stunned Kemp’s rid-
ley sea turtles in Cape Cod Bay during the cold-stunning
season (P3DWTMP variable removed). The two splits on the
tree suggest Kemp’s will cold-stun on Cape Cod Bay beaches
between 9 November and 20 December. Overall misclassification
= 46.1%, category 0 misclassification = 59.1%, category 1
misclassification = 9.4%.

875STILL ET AL. — Climatic Factors and Cold-Stunned Ridleys in Cape Cod Bay
ally an 8-terminal node tree was selected only 2% less than
the 5, 6, and 7-terminal node trees (Fig. 4C). We chose to
describe the 7-terminal node tree in addition to the modal
tree due to the strong support given to the larger tree size by
the fifty 10-fold cross validations.
The 2-terminal node tree had an overall
misclassification rate of 51.5%. However, the
misclassification rate for detecting the presence of cold-
stunned loggerheads (1) was much lower at 5.1%. This
simple model split the data using average sea surface
temperature (P3DWTMP), indicating loggerheads cold-
stunned at water temperatures ≤ 9.0ºC (Fig. 8). The
average sea surface temperature (P3DWTMP) (100) fol-
lowed by the day of the cold-stunning season (DAY) (70)
and the average air temperature (P3DATMP) (56.9) had
the highest relative importance of the six variables in-
cluded in the model. Wind direction (P3DPDIR) (12.7)
had moderate importance while wind speed (P3DWSPD)
(2.0) and barometric pressure (P3DBAR) (0.8) had little
influence in the model (Table 1E).
The 7-terminal node tree contained four of the six
explanatory variables, with the average sea surface tempera-
ture (P3DWTMP) variable used repeatedly throughout the
model. Loggerheads were predominantly present in two of
the 7-terminal nodes (terminal nodes 4 and 6) (Fig. 9). As in
the 2-terminal node model, the first split was based on the
average sea surface temperature (P3DWTMP); indicating
loggerheads do not cold-stun until average sea surface
temperature falls below 9.0ºC. The additional 5-terminal
nodes explained a greater level of interaction between the
climatic variables, suggesting two possible cold-stunning
scenarios. Under the first scenario, loggerheads were likely
to cold-stun after 5 December 5 (DAY = 188) when average
sea surface temperatures (P3DWTMP) were between 7.1
and 9.0ºC, average wind speed (P3DWSPD) ≤ 7.6 m/s, and
average barometric pressure (P3DBAR) > 1015.9 mm (Fig.
9). The second scenario suggested loggerheads cold-stun
when the average sea surface temperature (P3DWTMP) was
between 5.6 and 9.0ºC, and when the average wind speed
(P3DWSPD) exceeded 7.6 m/s (Fig. 9).
The relative importance of the explanatory variables
used in this classification tree was dominated by the average
sea surface temperature (P3DWTMP) (100). Day of the
cold-stunning season (DAY) (70), and average air tempera-
Figure 7. 8-terminal node classification tree predicting the daily
presence/absence of juvenile cold-stunned Kemp’s ridley sea turtles
in Cape Cod Bay during the cold-stunning season (P3DWTMP
variable removed). The model suggests Kemp’s will be predomi-
nantly present in three of the eight terminal nodes (3, 5, and 7).
Overall misclassification = 38.2%, category 0 misclassification =
47.6%, category 1 misclassification = 11.5%.
Figure 8. 2-terminal node classification tree suggesting cold-
stunned loggerheads will be predominantly present (1) at sea
surface temperatures below 9.0ºC. Overall misclassification =
51.5%, category 0 misclassification = 56.1%, category 1
misclassification = 5.1%.
Figure 9. 7-terminal node classification tree predicting the daily
presence/absence of juvenile cold-stunned loggerhead sea turtles
in Cape Cod Bay during the cold-stunning season. The model
suggests loggerheads will be predominantly present in two of the
seven terminal nodes (4 and 6). Overall misclassification = 29.4%,
category 0 misclassification = 31.1%, category 1 misclassification
= 13.1%.

876 CHELONIAN CONSERVATION AND BIOLOGY, Volume 4, Number 4 – 2005
ture (P3DATMP) (61) also had relatively high importance
values. Average wind speed (30.1) followed by average
barometric pressure (10.5) and average prevailing wind
direction (5.2) were moderately important in the model
(Table 1F).
DISCUSSION
As in other studies (Witherington and Ehrhart, 1989;
Burke et al., 1991), the prevailing wind direction associated
with cold-front storms was the most important factor in
determining the beach recovery location of cold-stunned
turtles in Cape Cod Bay. Given the semi-enclosed orienta-
tion of Cape Cod, the prevailing wind direction determines
the section of shoreline where cold-stunned turtles will most
likely be recovered. In Long Island Sound, however, the
prevailing wind direction also can determine the overall
cold-stunning magnitude each year. Due to the east-west
orientation of Long Island, turtles can be swept out of Long
Island Sound, if the prevailing wind direction is oriented
easterly (Burke et al., 1991).
The daily presence/absence cold-stunning models for
Kemp’s and loggerheads had similar overall structure and
variable importance. However, for sea surface temperature
and the day of the cold-stunning season, the values that
determined the model splits varied. For example, sea
surface temperature, when included, had the highest
relative importance for both species models. Yet, the
Kemp’s models indicated they were more likely to cold-
stun on days when the sea surface temperature was
between 7.5 and 10.4ºC, while for loggerheads the mod-
els suggested cold-stunning prevalence at sea surface
temperatures between 7.1 and 9.0ºC and 5.6 to 7.1ºC,
depending on the wind speed.
The timing of cold-stunning during the season, also
important for both models, varied for both species, with the
models suggesting Kemp’s typically start cold-stunning
much earlier in the season (ca. 9 November), followed later
by loggerheads (ca. 5 December). This seasonality is most
likely due to the dissimilar size of Kemp’s and loggerheads
that cold-stun each year in Cape Cod Bay; the mean SCL of
Kemp’s was smaller (26.9 cm) than that of loggerheads
(52.5 cm). The larger size for loggerheads enhances their
thermoregulatory capabilities and enables them to withstand
colder sea surface temperatures for longer periods of time,
relative to the much smaller Kemp’s.
The misclassification rates for predicting the absence
(0) of cold-stunned sea turtles were relatively high for both
the Kemp’s and loggerhead models. We believe this is due
to years in the data when there were relatively few cold-
stunned turtles recovered in Cape Cod Bay. We suspect that
these years with low cold-stun turtle numbers represent
years when there were few juvenile sea turtles in New
England coastal waters (Still, 2003). Therefore, during years
with few turtles in the region, no turtles would be recovered
despite the occurrence of the oceanographic and climatic
thresholds that trigger cold-stunning.
These classification tree results support the different
cold-stunning patterns observed for Kemp’s ridley and log-
gerhead sea turtles in Cape Cod Bay. Early in the season,
Kemp’s cold-stunning was most affected by the drop in sea
surface temperatures. Once the sea surface temperatures fall
below 10.4ºC, and cold front storms pass through the region
(low barometric pressures) with strong winds ranging
from north to westerly directions, cold-stunned Kemp’s
begin to wash up onto windward facing beaches. Al-
though node 5 suggested cold-stunned Kemp’s would be
present at barometric pressures exceeding 1009.5 mm,
the mean barometric pressure during November and
December from 1984 to 2001 was 1016.7 mm (n = 1098,
range 994.0–1037.9).
Previous cold-stunning studies provided descriptions
of the critical water temperature thresholds, but provided
little description of the climatic factors associated with cold-
front storms other than prevailing wind direction (Schwartz,
1978; Witherington and Ehrhart, 1989; Burke et al., 1991;
Morreale et al., 1992; Moon et al., 1997). Schwartz (1978)
reported that Kemp’s, loggerheads, and greens exhibited
cold-stunned floating behaviors in outdoor holding tanks
at water temperatures between 9.0 and 13.0ºC, with
death occurring between 5.0 and 6.5ºC for all species. In
Long Island Sound, New York, Morreale et al. (1992)
reported water temperatures below 10ºC during the peak
cold-stunning period of each year from 1985 to 1987.
Moon et al. (1997) demonstrated that juvenile Kemp’s
ridley and green sea turtles, in controlled laboratory
conditions, can adjust to slowly dropping water tempera-
tures (5–6ºC over an 8-week period), down to 15ºC,
without showing signs of severe cold stunning.
Witherington and Ehrhart (1989) described a series of
cold-stunning events in Mosquito Lagoon, Florida, from
1977 to 1986, where 342 greens, 123 loggerheads, and 2
Kemp’s were recovered. These events were triggered by
the arrival of severe cold fronts followed by several days
of unusually cold weather. They reported early morning
water temperatures generally below 8ºC when turtles
were recovered. They also suggested prevailing wind
direction determined where the turtles would be found in
the lagoon. Burke et al. (1991) also concluded prevailing
wind direction was a dominant factor in determining the
magnitude of cold-stunning events in Long Island Sound,
New York, between 1985 and 1988.
Although this study enhances our understanding of the
local extent and temporal scale factors that contribute to
cold-stunning events in Cape Cod Bay, Massachusetts, it
also provides a means to enhance recovery efforts of cold-
stunned turtles by the Sea Turtle Stranding and Salvage
Network. By monitoring sea surface temperature, the calen-
dar date, wind speed and direction, and barometric pressure,
the STSSN will be better able to maximize their recovery
efforts during those days with the highest potential occur-
rence for cold-stunned turtles. This will help to mobilize the
necessary staff, thereby increasing the survival rates of
recovered turtles.

877STILL ET AL. — Climatic Factors and Cold-Stunned Ridleys in Cape Cod Bay
ACKNOWLEDGMENTS
We would like to thank the many UMASS faculty and
graduate students in the Wildlife and Fisheries Conservation
program for their help in exploring the fascinating world of
statistics, including; Michel Sutherland, Kevin McGarigal,
Sam Cushman, Maile Neel, Brad Compton, and Joanna
Grand. Cheryl Ryder of the National Marine Fisheries
Service and Tom French of the Massachusetts Division of
Fisheries and Wildlife provided invaluable support and
encouragement throughout the project. This work would not
have been possible without the efforts of the countless
volunteers and staff from the Massachusetts Audubon
Society’s Wellfleet Bay Wildlife Sanctuary, New England
Aquarium Rescue and Rehabilitation Department, and mem-
bers of the Cape Cod National Seashore involved in the Sea
Turtle Stranding and Salvage Network. Their dedication and
hard work have increased the survival chances of hundreds
of cold-stunned turtles. Additionally, we are thankful for the
financial support provided by the National Marine Fisheries
Service and the NOAA CMER Program.
LITERATURE CITED
BLEAKNEY, J.S. 1965. Reports of marine life from New England and
eastern Canada. Canadian Field-Naturalist 79:120-128.
BREIMAN, L., FRIEDMAN, J.H., OLSHEN, R.A., AND STONE, C.J. 1984.
Classification and Regression Trees. Boca Raton: CRC Press, 368 pp.
BURKE, V.J., STANDORA, E.A., AND MORREALE, S.J. 1991. Factors
affecting strandings of cold-stunned juvenile Kemp’s ridley and
loggerhead sea turtles in Long Island, New York. Copeia
1991(4):1136-1138.
CART. 1998. Salford Systems, San Diego, California.
DE’ATH, G. AND FABRICIUS, K.E. 2000. Classification and regression
trees: a powerful yet simple technique for ecological data analysis.
Ecology 81(11):3178-3192.
EPPERLY, S.P., BRAUN, J., AND CHESTER, A.J. 1995. Aerial surveys for
sea turtles in North Carolina inshore waters. Fishery Bulletin
93(2):254-261.
HENWOOD, T.A. AND OGREN, L.H. 1987. Distribution and migrations
of immature Kemp’s ridley turtles, Lepidochelys kempi, and green
turtles, Chelonia mydas, off Florida, Georgia and South Carolina
(USA). Northeast Gulf Science 9:153-159.
KEINATH, J.A. 1993. Movements and behavior of wild and head-
started sea turtles. Ph.D. Thesis, College of William and Mary,
Williamsburg, Virginia.
KOVACH, W.L. 1994. Oriana for Windows, version 1.06. Kovach
Computing Services, Wales, U.K.
LAZELL, J.D., JR. 1976. This Broken Archipelago: Cape Cod and the
Islands, Amphibians and Reptiles. New York: Demeter Press, 260 pp.
LAZELL, J.D., JR. 1980. New England waters: critical habitat for
marine turtles. Copeia 1980(2):290-295.
MOON, D.Y., MACKENZIE, D.S., AND OWENS, D.W. 1997. Simulated
hibernation of sea turtles in the laboratory: I. feeding, breathing
frequency, blood pH, and blood gases. Journal of Experimental
Zoology 278:372-380.
MORREALE, S.J. AND STANDORA, E.A. 1994. Occurrence, movement
and behavior of the Kemps’ ridley and other sea turtles in New
York waters. Final Report to NYS Dept. of Env. Conservation,
Return a Gift to Wildlife Program, April 1988 – March 1993.
MORREALE, S.J. AND STANDORA, E.A. 2005. Western North Atlantic
waters: crucial developmental habitat for Kemp’s ridley turtles.
Chelonian Conservation and Biology 4(4):xxx-xxx.
MORREALE, S.J., MEYLAN, A.B., SADOVE, S.S., AND STANDORA, E.A.
1992. Annual occurrence and winter mortality of marine turtles in
New York waters. Journal of Herpetology 26(3):301-308.
MUSICK, J.A., BARNARD, D., AND KEINATH, J.A. 1994. Aerial estimates
of seasonal distribution and abundance of sea turtles near the Cape
Hatteras faunal barrier. In: Schroeder, B.A. and Witherington,
B.E. (Compilers). Proceedings of the Thirteenth Annual Sympo-
sium on Sea Turtle Biology and Conservation. NOAA Technical
Memorandum NMFS-SEFSC-341, pp. 121-123.
SCHWARTZ, F.J. 1978. Behavioral and tolerance responses to cold
water temperatures by three species of sea turtles (Reptilia,
Cheloniidae) in North Carolina. In: Henderson, G.E. (Ed.). Pro-
ceedings of the Florida and Interregional Conference on Sea
Turtles. Florida Marine Research Publications 33:16-18.
SHOOP, C.R. AND KENNEY, R.D. 1992. Seasonal distributions and
abundances of loggerhead and leatherback sea turtles in waters of
northeastern United States. Herpetological Monographs 6:43-67.
SPOTILA, J.R., O’CONNOR, M.P., AND PALADINO, F.V. 1997. Thermal
biology. In: Lutz, P.L. and Musick, J.A. (Eds.). The Biology of Sea
Turtles. Boca Raton, FL: CRC Press, pp. 297-314.
STILL, B.M. 2003. Factors affecting juvenile sea turtle cold-stunning
in Cape Cod Bay, Massachusetts. M.S. Thesis, University of
Massachusetts, Amherst.
WITHERINGTON, B.E. AND EHRHART, L.M. 1989. Hypothermic stun-
ning and mortality of marine turtles in the Indian River Lagoon
System, Florida. Copeia 1989(3):696-703.
Received: 3 June 2003
Revised and Accepted: 18 December 2004