Agricultural and Forest Meteorology 239 (2017) 58–70

Contents lists available at ScienceDirect

Agricultural and Forest Meteorology

journal homepage: www.elsevier.com/locate/agrformet

Physical controls on frost events in the central Andes of Peru using

in situ observations and energy flux models
Miguel Saavedra a,b,∗ , Ken Takahashi a
a
Instituto Geofísico del Perú, Calle Badajoz 169, Mayorazgo IV Etapa, Ate, Lima 15012, Perú
b
Facultad de Ciencias Físicas, Universidad Nacional Mayor de San Marcos, Lima 15081, Perú

a r t i c l e i n f o a b s t r a c t

Article history: Radiative frosts are a major hazard to agriculture in the tropical Andes of Peru, but there are very few
Received 11 June 2016 studies of their physical controls. In this study we focus on identifying and approximately estimating the
Received in revised form 14 February 2017 effect that physical variables have on both the downward surface longwave flux (LW↓ ) and the minimum
Accepted 17 February 2017
temperature (Tmin ). Through a combination of case studies and statistical analysis of in situ data in the
IGP Huancayo Observatory, we found that low cloud cover (CC), surface specific humidity (q), and soil
Keywords:
moisture are key factors controlling the day-to-day variability of Tmin , which is more pronounced in the
Frost
dry/cool season. We found that all frost days had q < 7 g/kg in the dry season and q < 5 g/kg in the wet
Andes
Longwave radiation
season, although it should be emphasized that q covaries with CC and soil moisture.
Radiative transfer We successfully validated a numerical soil heat diffusion model with data from a field campaign in
Soil thermal conductivity July 2010 and we used it, together with a radiative transfer model, to estimate the sensitivities of Tmin
and LW↓ to atmospheric and soil variables. With these results we estimated the partial contributions of
these variables to the overall day-to-day variability in Tmin and LW↓ . We found that low cloud cover is the
dominant factor, although specific humidity has a comparable role in the wet season. Lack of information
on the cloud liquid water path is an important source of uncertainty. Enhanced soil moisture has a strong
mitigating effect on frosts, although strong variability of soil moisture in the wet season could contribute
substantially to the development of frosts.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction The strong relation between cloudiness and minimum temper-

The Mantaro valley in the Peruvian Andes is arguably the main
agricultural region in the Peruvian Andes and is the main provider
of produce of Lima, the Peruvian capital. However, the agriculture
is extensive, largely dependent on rainfall, and is particularly vul-
nerable to frosts, which can destroy the crops and are, therefore,
considered the most damaging among extreme hydrometeorolog-
ical events (Trasmonte et al., 2008; Trasmonte, 2009; Núñez et al.,
2012). As reported by Garcilaso de la Vega (1609) in his “Comentar-
ios Reales”, the Inca farmers predicted frosts when they observed
cloudless skies at night and produced smoke as a way to mitigate
the frost damage to the crops. These prediction and mitigation prac-
tices persist in rural communities in the Mantaro valley (Martínez
et al., 2012), where the absence of late afternoon cloudiness has
been verified to increase the probability of freezing temperatures
by more than four times relative to cloudy skies (Saavedra, 2012).
∗ Corresponding author.
E-mail address: miguel.saavedra@igp.gob.pe (M. Saavedra).
ature indicates that the net radiative loss in the nocturnal surface
energy balance is a key process for the occurrence of frosts in the
Andes (Lhomme et al., 2007; Sanabria, 2009) and different meth-
ods have been tested to mitigate the impacts on the crops (Morlon,
1991; Lhomme and Vacher, 2002). Sicart et al. (2010) indicate
that clouds increase downward longwave by up to 55% near the
Zongo glacier in Bolivia. Furthermore, low clouds are expected to
emit more downward longwave radiation (e.g. Geiger et al., 2003;
Dai et al., 1999). Using world-wide station data, Dai et al. (1999)
concluded that the land diurnal temperature range (DTR), i.e. the
difference between the maximum and minimum temperature, can
be reduced in 25–50% by the presence of clouds and that their
spatial distribution determines that of DTR. They found a strong
correlation between minimum temperature and longwave radia-
tion, and at the same time with specific humidity, but with cloud
cover the correlation was weak. On the other hand, Huang et al.
(2006) used models to estimate that a cloud radiative forcing of
8.5 W m−2 is associated with an increase of 0.68 ◦ C in nocturnal
temperature. Studies developed in the Andes of Peru are scant.
Villegas (1991) indicate that frost in Mantaro basin, in the Andes

http://dx.doi.org/10.1016/j.agrformet.2017.02.019
0168-1923/© 2017 Elsevier B.V. All rights reserved.
 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70
of Peru, is associated with clear skies and dry air. The occurrence
of frost in this region is mainly between April and August, and the
lowest temperatures during June and July (Instituto Geofísico del
Perú, 2005).
Sanabria (2009) made use of two models (Lhomme and Guilioni,
2004; Cellier, 1993) built taking into account physical processes
to model minimum temperatures. Sanabria (2009) used long-
wave radiation and empirical methods (Swinbank, 1963; Brutsaert,
1975), but based on their results, they recommended that mea-
surements be made with pyrgeometers. Sicart et al. (2010)
implemented a parameterization of this longwave radiation in
tropical mountains using surface air temperature and vapor pres-
sure, as well as solar radiation, to estimate the cloud effects.
Although clouds are arguably the primary control on longwave
radiation, the air humidity and temperature also play a role in the
clear-sky emission and are key parameters used in empirical mod-
els (e.g. Lhomme et al., 2007; Sicart et al., 2010). Another physical
variable that affects the minimum temperature is the soil thermal
conductivity, which allows thermal exchange between the surface
and the deeper layers, damping the temperature changes at the
surface, and increases with the soil moisture, which implies that
precipitation in previous days can prevent the occurrence of a frost
event (e.g. Geiger et al., 2003).
To better understand and quantify the roles of different atmo-
spheric variables that control the minimum temperature at a site
in the Mantaro valley in the central Andes of Peru, in this study
we used a combination of in situ observations with numerical 1D
models. We first provide a description and climatic characteriza-
tion of the region of study. Then, we present an empirical analysis
of the relationship between the minimum temperatures and vari-
ables such as specific humidity, cloud cover and precipitation using
long-term meteorological data. After this, we use a soil heat diffu-
sion model and an atmospheric radiative transfer model to model
the minimum temperatures at the surface and LW↓ , respectively,
validated with data from a field experiment. Experiments with the
models are then used to assess the sensitivity of the minimum
temperature and LW↓ to changes in other atmospheric variables.
2. Study area and data
All the measurements in this study were made in the Huancayo
Observatory of the Geophysical Institute of Peru (IGP Observatory,
12.04◦ S, 75.32◦ W, 3350 m.a.s.l.), which is surrounded by non-
irrigated agricultural land in the Mantaro valley, in the central
Andes of Peru, located 12 km from the city of Huancayo and 7 km
from the Mantaro river (Fig. 1), between the western Andes and the
Fig. 1. IGP Observatory and topography around this location. At the East are located Huancayo city and Huaytapallana cordillera.
59
Huaytapallana Cordillera to the east, between the Pacific ocean and
the Amazon. In this research, we used three kinds of data: synop-
tic meteorological data from the Huayao station (WMO Id 84630),
data from a collocated automatic weather station and data from a
field experiment in the same site. The soil consists of clay covered
by short grass (2–3 cm high), partly dry due to lack of irrigation
during the dry season, with a leaf area index on the order of 0.8.
Synoptic meteorological data (1973–2006) include daily mini-
mum (Tmin ) and maximum (Tmax ) temperature, and precipitation PP
(at 07 local time or LT), as well as air temperature, relative humidity,
air pressure and cloud cover observed at 07, 13 and 19 LT. These
variables are measured at 2 m height above the soil surface. Air
temperature, relative humidity and air pressure were used to cal-
culate specific humidity. In the subsequent analysis, the cloud cover
(CC) and specific humidity (q) are considered as the mean of the 19
LT and 07 LT (next day) measurements to give an estimate of the
nocturnal values. We computed the daily climatology for this data
using harmonic analysis, using the first six annual harmonics, for
the different measurement times. In order to quantify the variabil-
ity, we calculated the interquartile range (IQR), that measures the
difference between the 75 and 25 percentile and is a measure of
the spread of the distribution. Additionally, the 2, 10, 25, 50, 75, 90,
and 98 percentiles for Tmin , Tmax and q are calculated. In the case of
CC and PP, we computed the monthly frequency of five categories
for precipitation (0 or dry, 0–2, 2–5, 5–10, and 10–42 mm/day) and
five categories for cloud cover (0 or cloudless, 0.5–2, 2.5–4, 4.5–6
and 6.5–8 oktas).
The field experiment took place in July 15–18, 2010, correspond-
ing to the dry, cold season. We measured LW↓ with a Kipp & Zonen
CGR3 pyrgeometer mounted on a 6 m tower to prevent obstacles
in its hemispheric field of view. The measurements were instanta-
neous at approximately hourly intervals, as the pyrgeometer could
not be connected to a datalogger at that time. We also used data
from a collocated automatic weather station in this period, which
included 2 m air temperature and relative humidity, air pressure,
and wind speed measured at height of 10 m, recorded every 10 min.
The sensors were a Vaisala thermo-hygrometer HMP35C (accuracy:
±0.2 ◦ C, ±3%RH), Vaisala barometer PTB101B (±0.5 hPa) and Young
wind monitor 05103 (±0.3 m/s) respectively.
We measured temperature in the soil using seven RadioShack
63-1032 indoor/outdoor digital thermometers. With the outdoor
(external) sensor, the temperatures at 50, 30, 20, 10, 5, 2 cm below
the soil surface and at 0 cm were measured; while the indoor sen-
sors measured at 10, 20, 40, 60, 80, 100, and 140 cm above the
surface. The temperature at 0 cm was measured with the digital
thermometer placed on top of the soil surface unsheltered and
60 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70

(a) value (nearly 5 mm d−1 ) between January and February, after which
5 it decreases steadily towards May.
The seasonal pattern of the precipitation is also approximately
4 followed by low cloud cover (Fig. 2b) and specific humidity (Fig. 2c),
Rain (mm d )
-1

including the existence of the secondary peak around early October,

3
even more significant relative to the main peak than in the case of
precipitation and also found particularly pronounced in low cloud
2
cover at 07 LT, albeit earlier by 2–3 weeks. Cloudiness varies from
1 slightly above 2 oktas in the dry season to over 6 oktas (at 19 LT)
in the peak of the wet season (Fig. 2b), while specific humidity
0 varies from 5–5.5 g/kg to 8.5–9 g/kg (Fig. 2c), larger than the val-
(b) ues reported by Garreaud (1999) for the Altiplano region (3.5 and
7
6.8 g/kg), probably due to the somewhat lower altitude and the
Low cloud cover (oktas)

07 LT influence of moist air masses from the Amazon on this region (e.g.
6
Junquas et al., 2017).
13 LT From August to February, low cloud cover at 19 LT is larger than
5
at 07 LT and 13 LT, reaching a peak of around 6 oktas between
19 LT January and February, while in the dry season it is below 3 oktas
4
at the three observing times (Fig. 2.b). For specific humidity q, the
3 diurnal variability is generally weak and is more pronounced in
Jun–Aug and Sep–Nov (Fig. 2c). Its minimum occurs around 13–15
2 LT, counter to what would be expected for a valley wind circulation
(c) (Junquas et al., 2017). Furthermore, the hourly data from the auto-
9
matic station (not shown) indicate dominance of the semi-diurnal
07 LT
Specific hum. (g/kg)

cycle, resulting in a secondary minimum around 05–06 LT, some-

8 what before the 07 LT synoptic observation, so the mean of the 07
13 LT and 19 LT observations may somewhat overestimate the nocturnal
7 average.
19 LT
Tmin follows a similar seasonal pattern to the previous
variables, ranging from around 7 ◦ C in December–February to
6
0 ◦ C in June–July (Fig. 2d). However, Tmax has a less pro-
nounced seasonality, ranging between 18.5 ◦ C (February) and 21 ◦ C
5 (October–November), with local minima in February and June
(d)
(Fig. 2d), most likely due to the attenuation of the wet season solar
20 radiation peak by wet season clouds. The weak seasonality in Tmax
Temperature (°C)

Maximum suggests that Tmin is controlled by longwave radiation from clouds

15
10
Minimum
5 season, 75% of all days have non-zero rainfall, while 30% (15%) have
0 sider rainy days in the wet season, the conditional frequency of
J F M A M J J A S O
Fig. 2. Climatology (1973–2006) of daily (a) precipitation, (b) low cloud cover and (c)
specific humidity (at 07, 13 and 19 LT), and (d) maximum and minimum temperature
for the IGP observatory.
unburied, beneath the grass. The sensor has a diameter of 5 mm and
length of 2 cm, so it did not disturb the grass. These thermometers
were not recording, so instantaneous readings were taken jointly
with LW↓ . The thermometers were selected from a set of 40 after
intercalibration between themselves, showing errors <0.5 ◦ C rela-
tive to the average, even though their nominal accuracy is ±1.8 ◦ C.
3. Climatology in the IGP observatory
In this section we report the thirty-three year (1973–2006)
climatology for the IGP Observatory. The seasonal cycle of pre-
cipitation (Fig. 2a) is primarily uni-modal with a wet season from
September to March (austral spring-fall) and a dry season between
May and July (austral fall-winter), with values below 0.5 mm d−1 .
The onset of the rainy season starts in August and reaches an initial
peak in early October of around 2.3 mm d−1 , after which it decreases
slightly and starts increasing from the end of October to the peak
rather than by solar radiation (Sicart et al., 2010). In fact, Tmin shows
a hint of a warm peak in October that approximately coincides with
the peak in cloudiness (Fig. 2b).
Rainfall presents substantial variability. In the peak of the wet
rainfall greater than 5 (10) mm d−1 (see Fig. 3a). If we only con-
N D rainfall greater than 5 (10) mm d−1 is maximum in February, with
values of 42% (22%), compared with 25 and 30% (10 and 15%) in
September and March, respectively, which indicates that the wet
season peak has not only a larger number of rainy days but also
a higher proportion of intense rainfall events. In the dry season, in
June and July, less than 15% of the days have non-zero rainfall, most
of it below 2 mm d−1 (see Fig. 3a). Consistent with the above, the
frequency of nocturnal clear skies decreases from 50% in the dry
season (May–July) to a minimum of under 16% in January–March
in the wet season. On the other hand, frequency of low cloud cover
above 6.5 oktas increases from less than 15% in the dry season to
40% in the wet season (see Fig. 3b).
The variability of Tmin in the dry season (IQR = 4.3 ◦ C) is greater
than in the wet season (IQR = 2.5 ◦ C), but this seasonal change is
larger for the lowest percentiles. For instance, the 2% percentile
varies between −6.1 ◦ C in July to 2.7 ◦ C in February, while the cor-
responding values for the 98% percentile are 7 ◦ C and 9.5 ◦ C (Fig. 3d).
If we consider that frosts occur for Tmin below 0 ◦ C, then these are
seen to occur more than 50% of the time in June and July (more
than 25% between May and August; Fig. 3d). In the wet season,
when crops are more vulnerable to frosts, these are much rarer.
The variability of specific humidity shows a somewhat simi-
lar seasonal pattern to Tmin , with IQR decreasing from 1.8 g/kg in
 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70 61

(a)

Cloud type
mm/day _ Cs _ Ci Cs _ Cs _ Cs Cs Cs Cs Cs Cs _ (a)

cover
2 6 1 8 6 1 6 7 3 3
0 0-2 2-5 5-10 10-42
Rain frequency (%)
_ _ _ _ _ _ _ As Ac _ _ _ _ As _
100 8 2 1
Sc Sc Sc St Cu Sc _ _ _ _ _ Cu _ Cu Sc
80 8 6 8 2 1 8 1 1 8

q at 2m (g/kg)
Temp. at 2m (°C)
60 20 10
15 8
40
10 6
20 5 4
0 0 1.4 °C 2
(b) January 17, 2005
oktas 14 15 16 17 18
Cover frequency (%)

0 0.5 - 2 2.5 - 4 4.5 - 6 6.5 - 8

100

Cloud type
80
Cs Cs _ Ci Cs Cs Cs Cs Cs Cs x Cs Ci Ci Ci (b)

cover
4 7 8 8 8 3 7 5 1 x 3 8 7 8
Ac _ _ _ _ _ _ _ _ _ x _ _ _ _
2 x
60 Sc Cb Sc _ _ _ _ Cu Sc _ x _ _ Cu _
2 1 8 1 2 x 1
40

q at 2m (g/kg)
Temp. at 2m (°C)
20 10
20
15 8
0 10 6
(c)
11 5 4
specific humidity (g/kg)

10 0 0.8 °C 2
9 November 23, 2005
98
8 20 21 22 23 24
90
7
75

Cloud type
6 Cs Cs _ Cs Cs _ Cs Cs Cs Cs _ Sc _ _ _ (c)
50

cover
1 3 6 4 1 6 1 3 1
5 Ac Ac _ Ac Ac _ _ _ _ _ Cs _ As _ _
25
4 10 7 1 1 1 1 4
_ Cu Sc Sc Cu Sc _ Cu _ _ _ _ Sc Sc Sc
3 2 3 8 1 2 8 1 4 8 8

q at 2m (g/kg)
Temp. at 2m (°C)
2 20 10
(d)
10 15 8
Minimum temp. (°C)

8 98 10 6
6 90 5 4
4
75 0 0.4 °C 2
2 October 15, 2006
0 50 12 13 14 15 16
-2 25
10
Cloud type

-4 2 _ _ Ci _ Cs Cs _ Cs Cs Cs Cs Cs Cs _ _ (d)
cover

7 4 3 8 6 7 7 4 2
-6 Ac _ _ _ _ _ As _ _ _ _ Ac _ _ _
8 8 1
(e) _ Ns St St Cu Sc _ _ _ _ _ Cu _ Cu Sc
8 1 8 4 2 1 3 8
24

q at 2m (g/kg)
Temp. at 2m (°C)
Maximum temp. (°C)

98 20 10
22 90
15 8
75
20 50 10 6
25 5 4
18 10
0
February 17, 2007 -0.3 °C 2
16 2
14 14 15 16 17 18
Time (days)
J F M A M J J A S O N D
Fig. 4. Hourly air temperature (◦ C; thick, red line) and specific humidity (g/kg; thin,
Fig. 3. Monthly frequency of occurrence (%) according to ranges in (a) precipi- blue line) at 2 m from the automatic weather station, and cloud cover (oktas) and
tation, and (b) low cloud cover (19 and 07 LT mean). Monthly percentiles of (c) cloud type observations at synoptic times in the IGP Observatory during selected
specific humidity (19 and 07 LT mean), (d) minimum temperature and (e) maxi- frost events in the Mantaro basin: (a) January 17, 2005, (b) November 23, 2005, (c)
mum temperature. The data was collected in the IGP observatory in the 1973–2006 October 15, 2006, and (d) February 17, 2007. (For interpretation of the references to
period. color in this figure legend, the reader is referred to the web version of the article.)

the dry season to 1.1 g/kg in the wet season, although it does not
present a marked asymmetry between the high and low percentiles
(Fig. 3c). Conversely, the variability of Tmax increases from the dry
to the wet season, approximately doubling the IQR to 3 ◦ C (Fig. 3e).
4. Observational analysis of frost events
4.1. Case studies
We have selected four cases of frost events that produced impor-
tant damage to crops in the region: January 17, 2005, November 23,
2005, October 15, 2006, and February 17, 2007 (Trasmonte, 2009).
These occurred within the frost-unfavorable wet and warm season,
between October and February, and therefore they were excep-
tional events. In order to learn what the conditions were before
and during the frosts, we analyzed the specific humidity and air
temperature at 2 m, cloud type and cloud cover, and precipitation
on the three days prior to each event.
Tmin in the selected days ranged from −0.3 ◦ C to 1.4 ◦ C (Fig. 4).
During each of those nights, specific humidity was below 5 g/kg,
in the 2% percentile (Fig. 3c), in all cases except January 17, 2005,
which had under 7 g/kg (10% percentile) but also the highest Tmin
(Fig. 4).
62 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70

Table 1 stronger in the dry season (correlation coefficient r = 0.83) than in

Daily precipitation (mm) in the days preceding the selected frost events.
January, 2005 12 13 14
Precip. (mm) 1.02 0 20.07
November, 2005 18 19 20
Precip. (mm) 0 2.54 0.51
October, 2006 10 11 12
Precip. (mm) 0.25 0 0
February, 2007 12 13 14
Precip. (mm) 4.06 3.05 7.11
Another common factor is the lack of low cloud cover those
nights (between 19 LT and 07 LT; Fig. 4), except for November 23,
2005, which only had 2 oktas (stratocumulus) at 19 LT (Fig. 4b), con-
sistent with reduced nocturnal LW↓ . Additionally, mid-level clouds
were also not observed, except in the night of January 17, 2005
(only 2 oktas). On the other hand, substantial high cloud cover
(cirrostratus) was observed before, during and after these events,
confirming that high clouds are not relevant for the occurrence
of frosts. It should be noted that the morning before the frost in
November 23, 2005, low clouds were also absent but a frost did not
occur.
During the two days prior to each frost event, daily precipita-
tion was zero or near zero (0.25 mm d−1 in November 22, 2005;
Table 1), which is an indication of reduced soil moisture during the
frost events. The largest precipitation event (20 mm d−1 ) took place
three days before the frost in January 17, 2005, which probably
helped mitigating the frost. On the other hand, except for the event
in January 17, 2005, the maximum temperature in the day prior to
each frost did not show any indication of reduction, as would be
expected in the case of advective frosts.
In general, common conditions prior to these exceptional frost
events include reduced specific humidity (i.e. clear-sky emission),
absence of low clouds, and lack of precipitation (i.e. soil moisture)
in the preceding two days. All these factors were probably impor-
tant for the development of these frosts. In the following section
we investigate these relations statistically using the longer term
record.
4.2. Statistical predictors of Tmin
We now analyze statistical relationships between Tmin and
other variables from the station observations (1973–2006) in the
peak of the wet season (December–February) and the dry season
(May–July). All correlations shown are significant at the 95% level.
We find a positive linear relationship between the nocturnal (19
and 07 LT mean) specific humidity (q) with Tmin (Fig. 5), which is
the wet season (r = 0.6), probably due to the smaller variability in
15 16 the latter. However, since the variability in q in both seasons is most
0 0 likely due to the same mechanism, i.e. changes in inflow from the
21 22 Amazon, a single relationship applied to the two seasons combined
0 0.25 produces a better fit (Fig. 5a; r = 0.88). In the dry season, all days with
13 14 Tmin < 0 ◦ C had q < 7 g/kg, while all days with q < 4 g/kg had Tmin < 0 ◦ C.
0 0 In the wet season, very few days had Tmin < 0 ◦ C and all of these had
15 16
q < 5 g/kg (below the 2% percentile). According to the regressions
0.25 0 coefficients, a 1 g/kg decrease in q implies a 1.93 ◦ C decrease in Tmin
in the dry season, but only 1.03 ◦ C in the wet season. Although the
above provides a useful reference for prediction purposes, it should
be noted that from a physical perspective, q could be a proxy for a
combination of correlated factors, including enhanced cloud cover
and soil moisture, as discussed later.
The relationship between low level CC and Tmin is illustrated in
Figs. 6a and 7a, in which we show that Tmin increases as CC increases,
probably due to the increase of LW↓ emitted by the clouds. In the
dry (wet) season, clear nights have a median Tmin of around 6 ◦ C
(4 ◦ C) lower than in overcast conditions. However, there are few
clear night samples in the wet season, so there is a larger scatter
in the distribution (Fig. 6a). The correlation coefficients between
Tmin and CC are r = 0.69 and 0.48 for the dry and wet season, respec-
tively, which is weaker than the one found for q. Furthermore, CC
is also correlated with q (r = 0.63 and 0.40), so it is not straightfor-
ward to empirically separate their effects on Tmin . In fact, fitting a
multiple regression model using both q and CC as predictors of Tmin ,
results in only a slightly higher r2 (0.74 and 0.43 for the dry and wet
season, respectively) than using only q (0.69 and 0.36), albeit sub-
stantially higher than with CC (0.47 and 0.23). It should be noted
that the estimates of q and CC might both have errors, in the first
case associated with the measurements of air temperature and rel-
ative humidity, and in the second case due to the subjectivity of the
visual estimation. This leads to uncertainty in the relative strengths
of their correlations with Tmin .
As a proxy for soil moisture in a particular day, we consider
if each of the preceding two days had above-zero precipitation
(wet or “W”) or not (dry or “D”) and we can classify each case
into four categories coded as DD, WD, DW, and WW (in the case
of WD, D is for the day before and W for the day before that), in
presumed increasing order of soil moisture. As expected, we find
that Tmin increases with soil moisture in both wet and dry seasons
(Figs. 6b and 7b), although with a larger median range in the dry
season (around 5 ◦ C) than in the wet one (2 ◦ C). However, since
precipitation is related to cloudiness, this could in part reflect the
influence of the latter, so we separate these effects by looking at
the cloud effects for the two extreme soil moisture conditions (DD
and WW).

(a) (Wet + Dry) season (b) Wet season (c) Dry season
12
r = 0.88 r = 0.60 r = 0.83
8
Tmin (°C)

4
0
-4
slope = 1.92 °C/(g/kg) slope = 1.03 °C/(g/kg) slope = 1.93 °C/(g/kg)
-8
0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
q (g/kg)
Fig. 5. Tmin versus q (19 and 07 LT mean) for wet+dry season (a), wet season (b) and dry season (c), including linear correlation and regression fits with 98% confidence
intervals.
 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70 63

10
(a) 8
8 (a)
6 4
4
All cases 0
2
All cases
0 -4
0 0.5-2 2.5-4 4.5-6 6.5-8
10 0 0.5-2 2.5-4 4.5-6 6.5-8
(b) 8
8 (b)
Minimum temperature (°C)

6 4

Minimum temperature (°C)

4
0
2
0 -4
DD WD DW WW
10 DD WD DW WW
(c) 8
8 (c)
6 4
4
DD cases 0
2
DD cases
0 -4
0 0.5-2 2.5-4 4.5-6 6.5-8
10 0 0.5-2 2.5-4 4.5-6 6.5-8
(d) Low cloud cover (oktas) 8
8 (d)
6 4
4
WW cases 0
2
WW cases
0 -4
0 0.5-2 2.5-4 4.5-6 6.5-8
0 0.5-2 2.5-4 4.5-6 6.5-8
Low cloud cover (oktas)
Low cloud cover (oktas)
Fig. 6. Minimum temperature under different precipitation and cloudy conditions
in the wet season. (a) Minimum temperatures for different cloud cover cases. (b) Fig. 7. The same as Fig. 6 but for the dry season.
The first character of each XX label is referred to two days before and the second
the day before, D means the daily precipitation is zero and W means the day was
rainy. (c) Minimum temperature according different cloud cover conditions when Table 2
there is no precipitation in the preceding two days (DD). (d) Minimum temperature Thermal conductivity (k) and thermal capacity (c) for clay soil and under different
according different cloud cover conditions when there is precipitation in the pre- conditions of soil moisture ().
ceding two days (WW). The minimum and maximum of the box represent the 25
 (% volume) 0 10 20 30 40
and 75 percentile, the center of the dark circle represent the median and the limits
of the arms of the boxes the 10 and 90 percentile. −1 −1
k (W m K ) 0.25 1.0 1.5 1.68 1.8
c (106 J m−3 K−1 ) 1.25 1.67 2.09 2.51 2.93

From Geiger et al. (2003).

We find that, in both seasons, the cloud effects on Tmin are
stronger in the DD (Figs. 6c and 7c) than in the WW condition
Figs. 6d and 7d, reflecting the mitigating effect of soil moisture to
that of CC variability. In clear nights, Tmin is lower by 2–3 ◦ C in the
DD cases relative to WW (Figs. 6c,d and 7c,d), although it should be
noted that in both seasons there are relatively few clear nights in
the WW condition.
4.3. Field campaign and LW↓ measurements
The field campaign provided us with more detailed informa-
tion on the processes affecting Tmin . It lasted three nights in the
dry season, which we label as Night I (July 15–16, 2010), Night
II (July 16–17, 2010), and Night III (July 17–18, 2010). In addi-
tion to the data described in Section 2, the upward longwave
flux (LW↑ ) and net upward diffusive energy flux at the soil sur-
face (Q0cm ) during the three nights were estimated assuming a
thermal conductivity (k) of 0.4 W m−1 K−1 and thermal capacity
(c) of 1.33 × 10−6 J m−3 K−1 . Considering these two last values
and using linear interpolation from the Table 2, we estimate
a soil moisture of 2%. While LW↑ was calculated using the
Stefan-Boltzmann law and an emissivity ε = 0.98, and Q0 cm was
estimated as a residual from the near-surface soil heat budget
(Q0 cm = Q3.5 cm − c∂[T]/∂t(0 − 3.5 cm), where Q3.5 cm is the energy
flux at 3.5 cm depth estimated by finite differencing between 2
and 5 cm, and c[T] is the heat content in the 0–3.5 cm layer).
For Night I, the temperature at 0 cm before 22 LT, was estimated
by extrapolating the 2 cm values using the 0–2 cm gradients from
the corresponding times in Night III. We were fortunate that the
three nights had distinct conditions, reflected in the respective 2 m
(0 cm) minimum temperatures of −1.9 ◦ C (−6.4 ◦ C), 5.6 ◦ C (1.2 ◦ C),
and 0.3 ◦ C (−3.4 ◦ C; Fig. 8a,b). Night II had full low cloud (stratocu-
mulus) cover, rare for July (Fig. 3b) in contrast with Nights I and III
(Fig. 8a). On the other hand, the specific humidity increased from
around 3 g/kg in Night I to around 5–6 g/kg in Nights II and III. In all
cases, the nocturnal (after 22 LT) 10 m wind speed was below 1 m/s,
the strongest diurnal wind (around 3 m/s) occurred before Night II.
The only precipitation in that month was on July 9 (5 mm), so the
64 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70

(a)
25
Night I Night II Night III 350

LW
↓

[ W m-2 ]
20 300

↓
[ m/s - g/kg - °C ]

↑
15 LW↑ 250

10 200

↓
Temp. at 2m
6

↓
Specific hum. at 2m
3

↓
Wind speed
0
-3 Cs As Cs Sc Sc Sc Sc -- Cs
6 1 7 5 8 8 3 0 7

(b) Time (day hour)

0 cm
2 cm
30 5 cm
Temperature (°C)

10 cm
20 cm
20 30 cm
50 cm

10

-10
15 12 16 00 16 12 17 00 17 12 18 00 18 12
Local time (day hour)

Fig. 8. (a) Temperature and humidity at 2 m height, wind speed at 10 m, cloud cover, estimated LW↑ and LW↓ recorded during campaign from 15 to 18 July, 2010 (19–07 LT
is shaded). (b) The temperatures on the soil surface and at 2, 5, 10, 20, 30 and 50 cm below (the completed 0 cm temperature for Night I are indicated with dots).

day-to-day changes  are probably not relevant in the comparative Table 3

Nocturnal near-surface energy budget terms (in W m−2 ) during the field campaign.
analysis of the three nights.
LW↓ , LW↑ , Q3.5 cm and Q0 cm .
To describe the soil thermal wave propagation, we take the
mean 50 cm soil temperature (14.8 ◦ C) as reference, as it pre- Energy fluxes Night I Night II Night III
sented little variation throughout the campaign (14.5–15.7 ◦ C), and LW↓ 239 300 252
express temperatures as anomalies relative to this reference. Thus, LW↑ 302 336 314
around 06 LT at the end of Nights I and III, the surface (0 cm) LW↑ − LW↓ 63 36 62
Q3.5 cm 51 26 47
temperature anomaly was −21.2 ◦ C and −18.2 ◦ C, respectively, and
c∂[T]/∂t (0–3.5 cm) −14 −5 −13
this cold signal progressed downward so that, at 10 cm depth, the Q0 cm 65 31 60
anomaly peaked around 10 LT with an amplitude of around 22% Q0 cm − (LW↑ − LW↓ ) +2 −5 −2
relative to the surface (Fig. 8b). During Night II, the variability in
LW↓ led to a double minimum, with the largest around 02 LT and
an anomaly of 13.5 ◦ C at the surface. At 5 cm depth, this complex vertical distribution. Thus, for example the  = 2% estimated from
structure disappeared while at 10 cm the peak anomaly was also linear interpolation from Table 2, is slightly smaller than the values
observed around 10 LT with a relative amplitude of 25% (Fig. 8b). of 2.7–5.3% estimated from Metop ASCAT Soil Moisture DataViewer
The above suggests a vertical e-folding scale and propagation speed (2017). In addition, Lu et al. (2007, 2009) found that k is almost
of the thermal wave of around 8 cm and 2.5 cm/h, respectively, con- insensitive to moisture in nearly dry clay soils with moistures as
sistent with the analytical solution for homogeneous soil under mentioned before; and increases as  exceeds 6% approximately,
diurnal surface forcing using parameters corresponding to dry clay so in our case this implies  = 8%. It is also possible that downward
soil (see Section 5.1). sensible heat flux played a role (Grant, 1974; André and Mahrt,
With respect to (LW↓ ), it initially decreased throughout Night I, 2000; Sun et al., 2003), but in a minor contribution due to the sta-
resulting in a nocturnal (19–07 LT) mean of 245 W m−2 . In Night II ble conditions and weak wind (see next paragraph). We note that
LW↓ increased, having a mean value of 300 W m−2 , probably asso- the Q1 cm (not shown) estimated from the 0 and 2 cm temperatures
ciated with the increase in q and CC, with some hourly variability was implausibly higher than the net radiative flux, indicating that
on the order of 50 W m−2 (Fig. 8a). The following day and through- our measurement at 0 cm, which corresponds to the grass canopy
out Night III, LW↓ decreased again and had a nocturnal mean of and not the soil surface, was inadequate for estimating the diffusive
255 W m−2 , associated with the reduced cloudiness, since q stayed heat flux, although it appears to have been adequate for estimating
relatively similar to Night II (Fig. 8a). Since air temperature was LW↑ .
higher in Night II, it could have also contributed to increasing LW↓ . Consistent with the surface cooling and weak nocturnal wind,
We estimate the mean net upward radiative flux for Nights I, II, and the atmospheric surface layer developed a very strong shallow
III as 63, 36, and 62 W m−2 , respectively (Table 3). inversion (below 20 cm altitude) (Fig. 9), with gradients of up to
The mismatch between the net longwave flux and Q0 cm indi- 16, 10 and 16 ◦ C/m in Nights I, II, and III, respectively (the gradi-
cates a downward surface energy flux on the order of 5 W m−2 ent for Night I was probably larger before 23 LT, but the necessary
(Table 3), which could be partly due to errors in the measure- data is not available). Between 20 and 140 cm above the surface,
ments of surface temperature by digital thermometer and lack the inversion was also strong (between 1 and 5 ◦ C/m) but not as
of accurate information on the local soil parameters and their much as at the surface.
 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70 65

140 Section 4.3). The model (Eqs. (1) and (2)) was discretized explicitly
(a)
with a time step t = 60 s and a vertical grid spacing of z = 1 cm.
100 We should emphasize that this model does not include grass cover.

Night I
The inclusion of this or plant canopy would be expected to lead
80
to lower skin temperature and higher soil surface temperature at
60 night (Deardorff, 1978; Herb et al., 2008).
40 The model also can be solved analytically, assuming LW↓ as a
20 periodic forcing with angular frequency ω, for the soil temperature:
0 F0
T (z, t) =  ez/d ei(z+vt)/d e−i ,
Height above the surface (cm)

-8 -6 -4 -2 0 2 4 6 8 10 12 (3)
140 2
(b) ( +
) + 2

where the e-folding scale is d = 2k/(cω), the downward prop-
100 

Night II
19h 02h
80 agation speed of the the thermal wave is v = 2kω/(c), the
22h 04h effective surfacedamping coefficient associated with soil heat
60
00h 06h transfer is = kcω/2, the phase-shift between the forcing
40
and the surface T is given by tan = /( +
) and
= 4ε(T0 )3 ,
20
where T0 is the reference surface temperature around which the
0 Stefan–Boltzmann law is linearized.
-8 -6 -4 -2 0 2 4 6 8 10 12
140 For the simulation of the three nights in the field campaign
(c)
(Section 4.3), we used homogeneous soil parameters corre-
100 Night III sponding to nearly dry clay (i.e.  = 2%, k = 0.40 W m−1 K−1 and
c = 1.33 × 106 J m−3 K−1 , based on Table 2), as well as a surface
80
emissivity ε = 0.98 corresponding to grass (Wilber et al., 1999;
60 Snyder et al., 1998). With these values, d = 7.4 cm and v = 1.9 cm/h,
40 similar to that observed in the field campaign (Section 4.3). The sim-
20 ulations for each night ran from 19 to 06 LT, using measured LW↓
0 as forcing and the observed ground temperature profiles at 19 LT as
-8 -6 -4 -2 0 2 4 6 8 10 12 initial conditions (Fig. 8). At the bottom (z =−50 cm), the tempera-
Temperature (oC) ture was set to the initial temperature at 19 LT, these were 14.8 ◦ C
for the first night and 14.7 ◦ C for the others. The ground temper-
Fig. 9. Nocturnal air temperature profiles above the soil surface at different times atures were generally well simulated, reproducing adequately the
during Nights (a) I, (b) II, and (c) III.
differences between the three nights and the variability in Night II
(Fig. 10), which confirms that the variations in LW↓ were the dom-
5. Numerical modeling of the sensitivity of Tmin and LW↓ inant factors controlling the surface temperature. However, there
was a negative bias that developed particularly through Nights I
5.1. Heat diffusion/energy balance modeling of Tmin and III of up to 3 ◦ C. This could be partly due to the neglect of the
small downward sensible heat flux (order of 5 W m−2 ; Sun et al.,
We have argued in Section 4.3 that the amplitude and propa- 2003).
gation of the nocturnal thermal wave matches well the theoretical We now test the effect that different factors have on Tmin
prediction of a diurnally-forced model of heat diffusion in homoge- through sensitivity experiments. The control model configura-
neous soil. Here we implement a numerical version of the ground tion had the initial ground temperature profile set to 15 ◦ C, with
heat diffusion model in which the surface energy balance is given the bottom temperature fixed at Tb ≡ T(z = −50 cm) = 15 ◦ C, k and
by the forcing by LW↓ , the response of LW↑ and heat diffusion into c for  = 0% and  = 30%, and surface emissivity ε = 0.98, forced
the ground. We then test the sensitivity of Tmin to changes in LW↓ , by a constant LW↓ = 280 W m−2 (the initial radiative imbalance is
initial conditions in temperature, , and soil emissivity. 111 W m−2 ). All runs start at 19 LT and end at 06 LT. For the sensi-
The heat-diffusion equation for the ground below the surface tivity experiments, we hold the control parameters fixed except for
(z ≤ 0) is: the one that is varied. We quantify the sensitivity of Tmin to a generic
  factor x, keeping everything else fixed, with the partial derivative
∂T ∂ ∂T
(c) =− −k (1) ∂Tmin /∂x.
∂t ∂z ∂z
The response of Tmin to LW↓ is close to linear between 200
where T is the soil temperature. In the subsequent section the val- and 350 W m−2 , with a mean sensitivity of 16 × 10−2 ◦ C/(W m−2 )
ues of Table 2 is used. (Fig. 11a). For the night-to-night fluctuations of around 50 W m−2
At the surface, we assume that shortwave (solar) radiation and we observed in the field campaign, this indicates variations in Tmin
the turbulent latent and sensible heat fluxes are negligible during of 8 ◦ C, although, in the Bolivian Andes, the variability in LW↓ in
the night, which is an adequate approximation for the conditions the dry season is up to 100 W m−2 and somewhat smaller in the
of the field campaign (Section 4.3). Thus, the boundary condition at wet season (Sicart et al., 2010). If  = 30%, the mean sensitivity is
the surface (z = 0) is: 8 × 10−2 ◦ C/(W m−2 ), half as in the dry case. Thus, changes in LW↓
can be expected to have more impact on Tmin in the dry season than
∂T  
−k = ε LW↓ − T 4 (2) in the wet one and could partially explain the larger Tmin variability
∂z in the former (Fig. 3d).
where the left side of the equation indicates the ground heat flux, We assess the effect of  as it varies between 0 and 40% by
LW↓ is imposed as a forcing from the field campaign or using ide- considering its effect on k and c (Tab. 2). The results indicate
alized values, and LW↑ is estimated with the Stefan–Boltzmann that Tmin increases with  and the sensitivity is largest for dry soil
law using a surface emissivity ε. The other boundary condition conditions, exceeding 0.6 ◦ C/% for moisture under 2% (Fig. 11b).
is implemented by prescribing T at z =−50 cm as a constant (see Therefore, increasing  from 0% to 5% increases Tmin by 3.2 ◦ C.
66 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70
20 20
Night I (a)
15 15
Temperature (°C)
10 10
5 5
0 0
-5 -5
-10 -10
19 20 21 22 23 00 01 02 03 04 05 06 19 20 21 22 23 00 01 02 03 04 05 06 19 20 21 22 23 00 01 02 03 04 05 06
Local Time
Fig. 10. Modeled (dashed lines) and observed (solid lines with dots) data used for validation of the model at the surface and 10 cm below. Triangles indicate values
reconstructed from observations.
However, the sensitivity is strongly reduced for higher , so an
increase from 20% to 40% would increase Tmin by 1.6 ◦ C. Although
the variability of  should be larger due to episodic precipitation in
the wet season (Fig. 3a), preliminary measurements of  in the IGP
Observatory (December, 2015–February, 2016) show typical vari-
ability in the 20–30% range, which by itself implies variability in
Tmin of the order of 1 ◦ C. However, in extreme cases, lack of precip-
itation for several days (e.g. Section 4.1) could alone theoretically
lower Tmin by up to 7.5 ◦ C (Fig. 11b).
Larger values of Tmax are usually an indication of consider-
able absorption of solar radiation in the soil surface. If this extra
heat could be stored long enough, it could perhaps mitigate the
subsequent frosts. Here we assess the sensitivity of Tmin to the
temperature at the beginning of the night (19 LT) by setting the
initial conditions that goes from an initial value of Ts at the sur-
face to 15 ◦ C at the bottom. For this we use an exponential profile
T(t = 0, z) = 15 ◦ C + (Ts − 15 ◦ C) exp(z/d), with d = 7.4 cm for the con-
trol parameters (Eq. (3)). We tested values of Ts ranging between
10 and 20 ◦ C and found that Tmin has very low sensitivity to the
initial surface temperature, with only 0.12 ◦ C increase in Tmin for
each 1 ◦ C increase in initial Ts (Fig. 11c). For  = 30%, this is further
reduced to 0.08 ◦ C. These results indicate strong damping, so that,
in these experiments, Ts approaches half-way towards its equilib-
rium value1 in 1 h, although the subsequent adjustment is slower
(after about 6 h, the departure from the equilibrium value is 20%
from the initial value).
The previous experiments tested the effect of a transient ini-
tial surface warming. Now we test the influence of heat flux from
the deep soil by keeping the initial Ts = 15 ◦ C while using different
values of the fixed bottom temperature Tb . Similarly to above, we
using the exponential profile T(t = 0, z) = Tb + (15 ◦ C − Tb ) exp(z/d),
with Tb ranging between 10 and 20 ◦ C. The results (Fig. 11d) indicate
a higher sensitivity than to the initial surface temperature: around
0.21 and 0.42 ◦ C increase in Tmin per degree increase in Tb for  = 0%
and  = 30%, respectively. Considering that the range between the
10 and 90 percentiles of Tb are 2 ◦ C and 1.3 ◦ C in the dry and wet sea-
son, respectively (not shown), and  = 30%, the associated variability
in Tmin is 0.8 and 0.5 ◦ C, respectively.
Experiments have shown that the surface emissivity in the
8–12 ␮m band increases around 0.06 as clay soils become wetter
(Mira et al., 2007; Sanchez et al., 2011). We tested the effect of vary-
ing the surface emissivity ε from 0.9 to 1.0 (not shown) and found
a modest overall decrease of Tmin of 1 ◦ C.
5.2. Radiative transfer modeling of LW↓
In the previous section we showed that LW↓ is the dominant
factor controlling the day-to-day variability in Tmin . Here we assess
1
In equilibrium, Ts ≈ (F − εTb4 )(kzb−1 + 4εTb3 )
−1
+ Tb , where Tb is the prescribed
temperature at z = zb .
20
Night II (b) Night III (c)
15
10
5
Obs. at 0 cm 0
Mod. at 0 cm
Obs. at 10 cm -5
Mod. at 10 cm
-10
Local Time Local Time
how LW↓ itself is affected by different atmospheric properties by
carrying out sensitivity experiments with the radiative transfer
model called Santa Barbara DISORT Atmospheric Radiative Transfer
(SBDART; Ricchiazzi et al., 1998). This model solves the radiative
transfer equation for a plane-parallel atmosphere using discrete
ordinate radiative transfer (DISORT) (Liou, 1973; Stamnes et al.,
1988). SBDART requires profiles of atmospheric pressure, temper-
ature, absolute humidity and ozone density as input and calculates
the vertical radiative fluxes as output. This model has previously
been tested to model longwave radiation in cloudless and overcast
conditions (Viúdez-Mora et al., 2009, 2015; Panicker et al., 2008).
To calculate the cloud effects, SBDART requires the following input
variables: cloud height and depth, cloud cover, effective cloud drop
radius (re ), and liquid water path (LWP, the mass of liquid water per
unit area in the atmospheric column).
Our control atmospheric profile has an idealized boundary layer
with depth zBL = 1200 m, within which the specific humidity (qa ) is
set to a constant value of q0 = 6 g/kg and the atmospheric tempera-
ture (Ta ) decreases from a surface value of T0a = 10 ◦ C with a lapse
rate 9.8 ◦ C/km. Above zBL , we used the mean July, 2010, Ta and qa
profiles from NCEP/NCAR Reanalysis, noting that Ta and qa are dis-
continuous at the boundary layer top. Pressure was estimated using
the hydrostatic equation and the ideal gas law using a surface pres-
sure of 690 hPa, consistent with the altitude of the IGP Observatory
(3330 m.a.s.l.). The ozone profile used is representative of tropical
regions McClatchey et al. (1972). Cloudiness was set to zero in the
control, but for the sensitivity experiments, we considered an ide-
alized homogeneous low cloud with a 200 m thickness and base
at 1000 m above the surface, with re fixed at 10 ␮m while LWP is
varied from 10 to 100 g/m2 and cloud cover between 0 and 8 oktas;
these parameters are based on radar and satellite data and global
observations (WMO, 1987; Minnis et al., 1992; Boers et al., 2000;
Pawlowska et al., 2000; Wang and Sassen, 2001; Gao et al., 2014).
We note that SBDART represents the effect of cloud cover by mod-
ifying the cloud optical depth by a factor of CC1.5 , with CC between
0 and 1. For the sensitivity experiments, T0a was varied between 4.5
and 15.5 ◦ C and q0 between 3 and 10 g/kg.
As q0 increases from 3 to 10 g/kg, LW↓ at the surface increases
approximately linearly, although the rate decreases from 6.3 to
4.5 W m−2 /(g/kg) (Fig. 12a; Ruckstuhl et al., 2007). Thus, in events
such as in November 23, 2005, and October 15, 2006 (Fig. 4b and c),
in which specific humidity decreased from around 7–8 to 3 g/kg at
a temperature around 10 ◦ C, this would result in a reduction in LW↓
of approximately 25 W m−2 . The sensitivity of LW↓ to q0 increases
with temperature, so that for q0 = 6 g/kg, it increases from 2.50 to
2.72 W m−2 /K as T0a increases from 5 to 15 ◦ C (Fig. 12a), making this
effect also significant in the wet/warm season.
LW↓ also increases approximately linearly with T0a , with sensi-
tivity rates increasing slightly from 2.3 to 2.8 W m−2 /K, as well as
increasing with q0 (Fig. 12b). However, the values should be taken
as an upper bound, as we are assuming a uniform Ta change in the
boundary layer, whereas nocturnal cooling is concentrated near the
 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70 67

(a) 15 (a)
280
10
260
5

LW↓ (W m-2)
Tmin (°C)

240
0

-5 220 Ta0 = 5 °C
θ= 0%
-10 Ta0 = 10 °C
θ = 30 % 200
a
T0 = 15 °C
-15
200 220 240 260 280 300 320 340 360 180
3 4 5 6 7 8 9 10
LW↓ ( W m-2 )
q0 (g/kg)
(b) 8 (b)
270
q0 = 3 g/kg
6 260
q0 = 6 g/kg
250

LW (W m-2)
q0 = 9 g/kg
Tmin (°C)

4
240
2 230

↓
0 220
210
-2
0 5 10 15 20 25 30 35 40 200
(c) 4 6 8 10 12 14 16
θ (%)
Ta0 (°C)
8 (c)
340
6 320
LW↓ (W m-2)
Tmin (°C)

4 θ= 0% 300

θ = 30 % 280
2
260 LWP = 10 g/m2
0
LWP = 50 g/m2
240
LWP = 100 g/m2
-2
10 12 14 16 18 20 220
0 1 2 3 4 5 6 7 8
Ts (°C)
Low cloud cover (oktas)
(d) 10
Fig. 12. (Top and middle) LW↓ sensitivity respect changes of temperature and spe-
8 cific humidity in the boundary layer. (Bottom) Sensitivity of LW↓ respect changes of
6 cloud cover with different LWP.
Tmin (°C)

4 θ= 0%
2 θ = 30 %
0 smaller sensitivity (around 25 W m−2 per okta) but extends to a CC
-2
-4
10 12 14 16 18 20
Tb (°C)
Fig. 11. Minimum temperature Tmin as a function of varying (a) LW↓ (W m−2 ), (b) soil
humidity (%), (c) initial surface temperature Ts (t = 0) (◦ C) and (d) bottom temperature
Tb (◦ C) in the sensitivity experiments.
surface. Since T0a is expected to also be affected by the changes in
LW↓ , which could be initially driven by changes in specific humid-
ity and clouds (see below), this could be considered at least partly
as a positive feedback that amplifies frosts.
On the other hand, although LW↓ increases monotonically with
cloud cover, how it does depends strongly on LWP (Fig. 12c). For
optically thin clouds (LWP = 10 g/m2 ), LW↓ is approximately linear
on CC, with a mean sensitivity of about 8 W m−2 per okta. However,
for optically thick clouds (LWP=100 g/m2 ), the sensitivity exceeds
40 W m−2 per okta at low CC (<1.5 oktas), but decreases by an
order of a magnitude for CC above 4 oktas (Fig. 12c). In the case
of intermediate thickness (LWP = 50 g/m2 ), the low CC regime has a
of around 2.2 oktas. For the thin clouds (LWP = 10 g/m2 ), a change
from clear sky to overcast sky results in an overall increase in LW↓
of 67 W m−2 , while for the thicker clouds (LWP = 50 and 100 g/m2 )
the increase is 95 W m−2 , similar to the results of Viúdez-Mora et al.
(2015) for a cloud base of 1000 m.
There are almost no estimates of LWP in the Andes. An annual
mean LWP in the Andes of Ecuador using NOAA-AVHRR (Advanced
Very High Resolution Radiometer) shows values between 75 and
300 g/m2 , suggesting that the thick cloud regime could be more
typical of clouds in the Andes, but this is primarily associated with
convective clouds (Bendix et al., 2005). More detailed measure-
ments are required, particularly for the low clouds that are more
relevant to frosts.
5.3. Relative contributions to the variability in minimum
temperature
In this section we try to assess the relative importance of the
contributions to the variability of Tmin by the different variables and
68 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70

Table 4 are smaller, resulting in a substantial but much smaller contribution

Estimated partial contribution to the variability of LW↓ and Tmin , for  = 0% (for  = 30%
to LW↓ and Tmin of around 11 W m−2 and 2 ◦ C, respectively, com-
in parenthesis). The observational value of ıTmin is also indicated.
pared to the dry season. In the case of thin clouds (LWP = 10 g/m2 )
x ıx Contribution Contribution the sensitivity varies less with CC and the differences in the con-
to ıLW↓ to ıTmin
tributions between the dry and wet are associated only with ıCC,
(W m−2 ) (◦ C)
potentially explaining 7 ◦ C and 4 ◦ C of ıTmin , respectively. Which
Dry season (MJJ) thickness regime is more appropriate for this region is unclear,
CC (thick) 5.5 okt 86 12.9 (6.9)
although thick clouds are probably more likely in the wet season,
CC (thin) ” 43 6.5 (3.7)
q0 3.6 g/kg 19.4 2.9 (1.6) when specific humidity is higher.
T0a −0.7 ◦ C −1.8 −0.3 (−0.1) Specific humidity variability ıq0 has a contribution to LW↓ and
Observed – – 10.8 ıTmin comparable to that of ıCC in the wet season (Table 4). In the
Wet season (DJF)
dry season, the contribution of ıq0 is larger than in the wet season,
CC (thick) 3.4 okt 11 1.7 (0.9)
CC (thin) ” 27 4.1 (2.2) partly due to the enhanced sensitivity at low q0 , but even so, it is
q0 2.3 g/kg 11.0 1.7 (0.9) substantially smaller than that of ıCC (Table 4). With respect to
T0a −1.5 ◦ C −3.9 −0.6 (−0.3) air temperature, our estimated effect of ıT0a is a small reduction of
Observed – – 6.1 ıTmin (Table 4). The reason for this opposing effect is that the clear
sky that favors low Tmin also results in high surface insolation that
increases the maximum temperature during the previous day. This
parameters by considering the sensitivities (Sections 5.1 and 5.2)
effect is minor relative to that of CC and q0 . In all of the previous
and their climatological ranges of variability (Section 3). Since the
cases (including CC), for wet soil ( = 30%), the contributions to ıTmin
primary driver of Tmin is the atmospheric longwave radiation, we
through LW↓ decrease by a half. This is probably more relevant in
first estimate the influence of the atmospheric variables x (q0 , T0a
the wet season but can be used as a mitigation measure for crops
or CC) on the variability of the latter (ıLW↓ ) as
in the dry season.
 ∂LW ↓ In general, low cloud cover appears to be the dominant factor
ıLW ↓ = ıx, (4) controlling the variability of Tmin , although in the wet season, the
∂x
x humidity also plays a comparable role, while  has a large mitigat-
ing effect. The total estimate of ıTmin (Eq. (5)) is 8.4–15.5 ◦ C and
where the contribution to ıLW↓ by each variable x is the product of
1.5–2.8 ◦ C in the dry and wet seasons, respectively. This is roughly
the corresponding sensitivity (∂LW↓ /∂x) and the variations ıx.
consistent with the observational values (Table 4), although it is
Similarly, the contributions to the variability of Tmin is estimated
somewhat overestimated for the dry season, for which thick clouds
from:
produce a large part of this, perhaps due to inadequacies in the
 ∂T treatment of clouds in SBDART, and underestimated for the wet
min
ıTmin = ıx, (5)
∂x season, but this could be enhanced by strong temporal variability
x
of  that, in the extreme case that it varies from 0 to 30%, it would
where in this case x stands for LW↓ , , Ts (t = 0), Tb and ε. result in a contribution to ıTmin of 7.6 ◦ C, which would take the total
To estimate the ıx, we composite the data according to whether estimate above the observed (10–12 ◦ C).
the 2 m minimum air temperature is below its 10 percentile (“cold”)
or above its 90 percentile (“warm”) and ıx is calculated as the aver-
age x of the warm group minus the average x of the cold group. 6. Conclusions
This is done separately for the dry and wet seasons, so we neglect
the contributions from Tb and ε, which vary little within each sea- Radiative frosts are one of the main hazards for agriculture in
son (see Section 5.1). Additionally, q0 is taken as the mean 2 m the tropical Andes of Peru but they are poorly understood from
specific humidity at 19 LT of the previous day and 07 LT of the a physical perspective. This study was focused on characterizing
current day, while T0a is estimated as the maximum 2 m air temper- and understanding the physical controls on the minimum temper-
ature of the previous day, as it is expected to be a better indicator ature in this region, combining case studies, field measurements,
of the boundary layer temperature above the surface layer due to statistical analysis, and theoretical and numerical modeling.
daytime convective mixing. The surface soil temperature at 19 LT Firstly, we characterized the climatology in the IGP Observatory
(Ts (t = 0)) is taken to be similar to air temperature at 2 m since, in the central Andes of Peru (3330 m.a.s.l.), focusing on air tem-
although the surface air at 19 LT can be more than 2 ◦ C colder than perature, specific humidity, low cloud cover, and precipitation. We
at 2 m (Fig. 9), its day-to-day variability is probably a reasonable found that the unimodal precipitation regime, with pronounced
proxy of ıTs (t = 0). Because we do not have long series of  mea- wet and dry seasons (peaks in December–February and May–July,
surements, we consider the range between dry ( = 0%) and wet respectively), is similar to the seasonal pattern of the specific
( = 30%) soil. Variability (ıx) of CC, q0 , T0a , and , their estimated humidity, low cloud cover and minimum temperature, while the
partial contribution to the variability of LW↓ [(∂LW↓ /∂x)ıx, only for maximum temperature has relatively weak seasonality. Since our
q0 , T0a , CC] and the estimated partial contribution to the variability interest is the extreme low temperature events, we also character-
of Tmin [(∂Tmin /∂LW↓ )(∂LW↓ /∂x)ıx for q0 , T0a , CC, and (∂Tmin /∂x)ıx ized the seasonality of the day-to-day variability. The variability of
for Ts (t = 0)], for  = 0% and  = 30% are shown in Table 4. the Tmin is more pronounced in the dry season and, considering 0 ◦ C
According to our calculations, the largest contribution to the as the threshold for frosts, we found that these occur over 50% of
variability of LW↓ and Tmin is by far due to cloud cover. For the case the time in June and July, while during the wet season, which is
of thick clouds (LWP = 100 g/m2 ), we find that the typical variabil- when crops are more vulnerable, frosts are much rarer.
ity (associated with the variability of Tmin ) in CC, i.e. ıCC, produces We analyzed four frost events that produced damage to crops
changes in LW↓ and Tmin of around 86 W m−2 and 13 ◦ C, respec- in the wet season and found common characteristics during those
tively, in the dry season (Table 4). This is partly due to the large events: absence of low cloud cover and reduced specific humidity
ıCC in this season but also to the high sensitivity of LW↓ in the during the nights of these events, and lack of rainfall in the pre-
low CC/high LWP regime (Section 5.2, Fig. 12c). The wet season is vious two days. These observations were then confirmed with a
cloudier and so both the variability ıCC and the sensitivity of LW↓ statistical analysis of long term data (1973–2006), that showed a
 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70
strong positive correlation between nocturnal q and (Tmin ), partic-
ularly in the dry season. Although, from a practical perspective, all
frost days had q < 7 g/kg and q < 5 g/kg in the dry and wet season
respectively. This relationship appears to reflect the action of mul-
tiple factors correlated with q, specially low cloud cover, which also
presents a substantial correlation with Tmin , while nights with clear
skies have around 6 ◦ C (4 ◦ C) lower Tmin than with overcast skies in
the dry (wet) season. We estimated the effect of soil moisture by
considering the precipitation in the two preceding days and found
that, in the dry (wet) season, the driest category had around 5 ◦ C
(2 ◦ C) lower Tmin than the wettest category.
Data from a three-day field experiment in July 2010 provided
longwave radiative fluxes, as well as near-surface air and soil tem-
perature profiles, in addition to conventional meteorology. The
variability of the soil temperature was consistent with the the-
oretical diurnal thermal wave propagation using soil parameters
for nearly dry clay soil. Using the downward longwave flux as
forcing, a numerical heat diffusion model was able to simulate
the observations well, indicating that indeed Tmin is largely con-
trolled by the longwave forcing and that the model parameters
were adequate, although the model presented a cold bias beneath
the surface, perhaps because the model parameters were assumed
vertically uniform. We should also note that although our results
appear adequate for grass-covered soil, different vegetation heights
or types will impact on the soil surface temperature, so further
studies should assess such effects in this region.
We performed experiments with the heat diffusion model to
estimate the sensitivity of Tmin to changes in the longwave forcing,
soil moisture, the initial temperature profile, and the bottom tem-
perature (50 cm depth). Similarly, we used the SBDART radiative
transfer model to estimate the sensitivity of the downward long-
wave flux to the atmospheric temperature and specific humidity,
as well as the low cloud cover. With these sensitivities, we pro-
duced an estimation of the partial contributions from the different
controls on both the longwave forcing and the minimum temper-
ature in this region. In general, low cloud cover appears to be the
dominant factor, although q plays a comparable role in the wet sea-
son, while soil moisture has a large mitigating effect. By adding the
partial contributions from these individual processes, our calcula-
tion is consistent with although somewhat overestimates the total
range of Tmin compared to the observed. Lack of information on the
actual LWP and its joint treatment with cloud cover in SBDART are
potential sources of error. Additionally, the assumption of linear-
ity implicit in the analysis is not strictly correct as we have shown
nonlinearities in the sensitivities.
This study has provided a needed documentation of some
aspects of the physical controls on minimum temperature in the
Andes. However, it has also highlighted some urgent needs, par-
ticularly the need to have more comprehensive and sustained
measurements of surface fluxes (radiative and turbulent) and soil
moisture and temperature, but also of boundary layer and cloud
properties, not only to understand the processes acting in the
present, but to be able to adequately validate and correct climate
change models in the Andean region.
Acknowledgements
This work was supported by the project “Enhancing Resilience
of Rural Communities to Drought, Floods and Frost in the Mantaro
Valley, Peru (MAREMEX Mantaro)” (IDRC 105567).
References
André, J.C., Mahrt, L., 2000. The nocturnal surface inversion and influence of
clear-air radiative cooling. J. Atmos. Sci. 39, 864–878.
69
Bendix, J., Rollenbeck, R., Gottlicher, D., Cermak, J., 2005. Cloud occurrence and
cloud properties in Ecuador. Clim. Res. 30 (2), 133.
Boers, R., Russchenberg, H., Erkelens, J., Venema, V., van Lammeren, A., Apituley, A.,
Jongen, S., 2000. Ground-based remote sensing of stratocumulus properties
during CLARA, 1996. J. Appl. Meteorol. 39 (2), 169–181.
Brutsaert, W., 1975. On a derivable formula for long-wave radiation from clear
skies. Water Resour. Res. 11 (5), 742–744.
Cellier, P., 1993. An operational model for predicting minimum temperatures near
the soil surface under clear sky conditions. J. Appl. Meteorol. 32 (5), 871–883.
Dai, A., Trenberth, K.E., Karl, T.R., 1999. Effects of clouds, soil moisture,
precipitation, and water vapor on diurnal temperature range. J. Clim. 12 (8),
2451–2473.
Deardorff, J.W., 1978. Efficient prediction of ground surface temperature and
moisture, with inclusion of a layer of vegetation. J. Geophys. Res. 83, 1889.
Gao, W., Sui, C.H., Hu, Z., 2014. A study of macrophysical and microphysical
properties of warm clouds over the Northern Hemisphere using
CloudSat/CALIPSO data. J. Geophys. Res.: Atmos. 119 (6), 3268–3280.
Garcilaso de la Vega, 1609. Libro VII, Capítulo V. Comentarios Reales de los Incas.
Colección de autores peruanos. Editorial Universo, Lima, Perú.
Garreaud, R., 1999. Multiscale analysis of the summertime precipitation over the
Central Andes. Mon. Weather Rev. 127, 901–921.
Geiger, R., Aron, R.H., Todhunter, P., 2003. The Climate Near the Ground, sixth
edition. Rowman and Littlefield Publishers, 584 pp.
Grant, A.I.M., 1974. An observational study of the evening transition
boundary-layer. Q. J. R. Meteorol. Soc. 123, 657–677.
Herb, W.R., Janke, B., Mohseni, O., Stefan, H.G., 2008. Ground surface temperature
simulation for different land covers. J. Hydrol. 356 (3), 327–343.
Huang, Y., Dickinson, R.E., Chameides, W.L., 2006. Impact of aerosol indirect effect
on surface temperature over East Asia. Proc. Natl. Acad. Sci. U. S. A. 103 (12),
4371–4376.
Instituto Geofísico del Perú, 2005. Diagnóstico de la Cuenca del Mantaro Bajo la
Visión del Cambio Climático. Fondo editorial del Concejo Nacional del
Ambiente, Lima-Perú, 90 pp.
Junquas, C., Takahashi, K., Condom, T., Espinoza, J.C., Chavez, S., Sicart, J.E., Lebel, T.,
2017. Understanding the influence of orography on the precipitation diurnal
cycle and the associated atmospheric processes in the central Andes. Clim.
Dyn. (submitted for publication).
Lhomme, J.P., Guilioni, L., 2004. A simple model for minimum crop temperature
forecasting during nocturnal cooling. Agric. Forest Meteorol. 123 (1), 55–68.
Lhomme, J.P., Vacher, J.J., Rocheteau, A., 2007. Estimating downward long-wave
radiation on the Andean Altiplano. Agric. Forest Meteorol. 145 (3), 139–148.
Lhomme, J.P., Vacher, J.J., 2002. Modelling nocturnal heat dynamics and frost
mitigation in Andean raised field systems? Agric. Forest Meteorol. 112 (3),
179–193.
Liou, K.N., 1973. A numerical experiment on Chandrasekhar’s discrete-ordinate
method for radiative transfer: applications to cloudy and hazy atmospheres. J.
Atmos. Sci. 30 (7), 1303–1326.
Lu, S., Ren, T., Gong, Y., Horton, R., 2007. An improved model for predicting soil
thermal conductivity from water content at room temperature. Soil Sci. Soc.
Am. J. 71, 8–14.
Lu, S., Ju, Z., Ren, T., Horton, R., 2009. A general approach to estimate soil water
content from thermal inertia. Agric. For. Meteorol. 149, 1693–1698.
Martínez, A.G., Núñez, E., Beraún, V., Enciso, L., Céspedes, L.,2012. Conocimiento
local sobre tiempo y clima en el valle del Mantaro. In: Manejo de riesgos de
desastres ante eventos meteorológicos extremos en el valle del Mantaro.
Instituto Geofísico del Perú, pp. 54–60.
McClatchey, R.A., Fenn, R.W., Selby, J.A., Volz, F.E., Garing, J.S., 1972. Optical
Properties of the Atmosphere (No. AFCRL-72-0497). Air Force Cambridge Res.
Lab., Hanscom.
Metop ASCAT Soil Moisture DataViewer, 2017. http://rs.geo.tuwien.ac.at/dv/ascat/.
Minnis, P., Heck, P.W., Young, D.F., Fairall, C.W., Snider, J.B., 1992. Stratocumulus
cloud properties derived from simultaneous satellite and island-based
instrumentation during FIRE. J. Appl. Meteorol. 31 (4), 317–339.
Mira, M., Valor, E., Boluda, R., Caselles, V., Coll, C., 2007. Influence of soil water
content on the thermal infrared emissivity of bare soils: implication for land
surface temperature determination. J. Geophys. Res.: Earth Surf. (2003–2012)
112 (F4).
Morlon, P., 1991. Variations climatiques et agriculture sur l’Altiplano du lac
Titicaca (Pérou-Bolivie): une approche préliminaire. La Métórologie 39, 10–29.
Núñez, E., Enciso, L., Céspedes, L., Martínez, A.G.,2012. Percepciones de la población
rural y urbana en el valle del Mantaro. In: Manejo de riesgos de desastres ante
eventos meteorológicos extremos en el valle del Mantaro. Instituto Geofísico
del Perú, pp. 42–45.
Panicker, A.S., Pandithurai, G., Safai, P.D., Kewat, S., 2008. Observations of enhanced
aerosol longwave radiative forcing over an urban environment. Geophys. Res.
Lett. 35 (4).
Pawlowska, H., Brenguier, J.L., Burnet, F., 2000. Microphysical properties of
stratocumulus clouds. Atmos. Res. 55 (1), 15–33.
Ricchiazzi, P., Yang, S., Gautier, C., Sowle, D., 1998. SBDART: a research and
teaching software tool for plane-parallel radiative transfer in the Earth’s
atmosphere. Bull. Am. Meteorol. Soc. 79 (10), 2101–2114.
Ruckstuhl, C., Philipona, R., Morland, J., Ohmura, A., 2007. Observed relationship
between surface specific humidity, integrated water vapor, and longwave
downward radiation at different altitudes. J. Geophys. Res.: Atmos.
(1984–2012) 112 (D3).
70 M. Saavedra, K. Takahashi / Agricultural and Forest Meteorology 239 (2017) 58–70
Saavedra, M.,2012. ¿Sirve el conocimiento tradicional desde un punto de vista
físico?: Estudio de caso sobre el pronóstico de heladas en el valle del Mantaro.
In: Manejo de riesgos de desastres ante eventos meteorológicos extremos en el
valle del Mantaro. Instituto Geofísico del Perú, pp. 61–63.
Sanabria, J., 2009. Calibración y validación de modelos de pronóstico de heladas en
el valle del Mantaro. Universidad Nacional Agraria La Molina.
Sanchez, J.M., French, A.N., Mira, M., Hunsaker, D.J., Thorp, K.R., Valor, E., Caselles,
V., 2011. Thermal infrared emissivity dependence on soil moisture in field
conditions. IEEE Trans. Geosci. Remote Sens. 49 (11), 4652–4659.
Sicart, J.E., Hock, R., Ribstein, P., Chazarin, J.P., 2010. Sky longwave radiation on
tropical Andean glaciers: parameterization and sensitivity to atmospheric
variables. J. Glaciol. 56 (199), 854–860.
Snyder, W.C., Wan, Z., Zhang, Y., Feng, Y.Z., 1998. Classification-based emissivity
for land surface temperature measurement from space. Int. J. Remote Sens. 19
(14), 2753–2774.
Stamnes, K., Tsay, S.C., Wiscombe, W., Jayaweera, K., 1988. Numerically stable
algorithm for discrete-ordinate-method radiative transfer in multiple
scattering and emitting layered media? Appl. Opt. 27 (12), 2502–2509.
Sun, J., Burns, S.P., Delany, A.C., Oncley, S.P., Horst, T.W., Lenschow, D.H., 2003. Heat
balance in the nocturnal boundary layer during CASES-99. J. Appl. Meteor. 42,
1649–1666.
Swinbank, W.C., 1963. Longwave radiation from clear skies. Q. J. R. Meteorol. Soc.
89 (381), 339–348.
Trasmonte, G., Chavez, R., Segura, B., Rosales, J.L., 2008. Frost risks in the Mantaro
river basin. Adv. Geosci. 14 (14), 265–270.
Trasmonte, G., 2009. Propuesta de gestión de riesgo de heladas, que afectan a la
agricultura del valle del Mantaro. Tesis para optar el grado académico de
Maestra en Ecología y Gestión Ambiental. Universidad Ricardo Palma.
Villegas, E., 1991. Zonificación del valle del Mantaro según la intensidad y riesgo de
ocurrencia de heladas radiacionales, Tesis para optar el título de Ingeniero
Meteorólogo. Universidad Nacional Agraria La Molina, 134 pp.
Viúdez-Mora, A., Calbó, J., González, J.A., Jiménez, M.A., 2009. Modeling
atmospheric longwave radiation at the surface under cloudless skies. J.
Geophys. Res.: Atmos. (1984–2012) 114 (D18).
Viúdez-Mora, A., Costa-Surós, M., Calbó, J., González, J.A., 2015. Modeling
atmospheric longwave radiation at the surface during overcast skies: the role
of cloud base height. J. Geophys. Res. Atmos. 120, 199–214.
World Meteorology Organizations, 1987. International Cloud Atlas, vol. II. World
Meteorology Organization, 206 pp.
Wang, Zhien, Sassen, Kenneth, 2001. Cloud type and macrophysical property
retrieval using multiple remote sensors. J. Appl. Meteor. 40, 1665–1682.
Wilber, A.C., Kratz, D.P., Gupta, S.K., 1999. Surface Emissivity Maps for Use in
Satellite Retrievals of Longwave Radiation. NASA Tech. Publ.
NASA/TP-1999-209362, 35 pp.