On 17 August 2007, the center of Hurricane Dean passed within 92 km of the mountainous island of Dominica in the West Indies. Despite its distance from the island and its category 1–2 state, Dean brought significant total precipitation exceeding 500 mm and caused numerous landslides. Four rain gauges, a Moderate Resolution Imaging Spectroradiometer (MODIS) image, and 5-min radar scans from Guadeloupe and Martinique are used to determine the storm’s structure and the mountains’ effect on precipitation. The encounter is best described in three phases: (i) an east-northeast dry flow with three isolated drifting cells; (ii) a brief passage of the narrow outer rainband; and (iii) an extended period with south-southeast airflow in a nearly stationary spiral rainband. In this final phase, from 1100 to 2400 UTC, heavy rainfall from the stationary rainband was doubled by orographic enhancement. This enhancement pushed the sloping soils past the landslide threshold. The enhancement was caused by a modified seeder–feeder accretion mechanism that created a “dipole” pattern of precipitation, including a dry zone over the ocean in the lee. In contrast to normal trade-wind conditions, no terrain triggering of convection was identified in the hurricane environment.
1. Hurricane passage
On Friday, 17 August 2007, as Hurricane Dean was gathering strength, it passed through the Lesser Antilles, bringing heavy rain and wind to Martinique and Dominica on its right flank. Although not as damaging as Hurricane David in 1979, Dean still caused extensive crop damage and landslides on Dominica. As part of an experiment on orographic precipitation in the tropics, the Yale Dominica Precipitation Project installed a line of four rain gauges across the southern mountains of Dominica. These instruments, together with Météo-France radars on Guadeloupe and Martinique, provide a rare opportunity to see how hurricane precipitation can be modified by steep terrain.
The structure of Dean is shown in the Moderate Resolution Imaging Spectroradiometer (MODIS) image from 1445 UTC (Fig. 1). At this moment, the storm center had already passed the longitude of Dominica (61.3°W). As the storm itself was strengthening and drifting away, the winds near Dominica were steady or weakening, and precipitation in Dominica was in its strongest phase. The structure of Dean’s upper clouds in Fig. 1 is very complicated, with two or three broad spiral bands. The large outflow shield completely obscures the island, and no evidence of the island’s disturbing effect is seen in the cirrus shield or convective tops.
We constructed an idealized model of Dean’s track by computing a linear regression of the center latitude and longitude versus time, using the National Hurricane Center (NHC) advisories in Table 1. This track is overlain on the MODIS image in Fig. 1. We also regressed the maximum wind speed versus time to obtain VMAX(t) = 0.693t + 41.9, where t is the time (UTC) on 17 August (e.g., at 1200 UTC, VMAX = 50.2 m s−1). According to these fits, Dean’s westward speed was approximately 31 km h−1 or 8.6 m s−1. Its closest approach to Dominica (15.3°N, 61.3°W) was 92 km at 1145 UTC on 17 August. At this time, Dominica went from Dean’s front right to its rear right quadrant. The storm intensified slightly as it passed, but its major intensification did not occur until late the following day when Dominica was well beyond Dean’s reach.
To characterize the local environment of Dominica within the hurricane, we constructed a simple model of the changing wind vector near Dominica by assuming that the wind field is the sum of a uniform environmental wind (VE = 8 m s−1 easterly) and an axisymmetric vortex,
with rM = 40 km and b = 0.5 (Chan and Williams 1989). The maximum tangential wind in (1) is given by VM(t) = VMAX(t) − VE. The result of the model is shown in Fig. 2, where it is compared against wind speeds obtained from three soundings (Table 3) and from hourly tracking of precipitation cells in the Guadeloupe radar. The agreement is satisfactory, considering that we have neglected the hurricane beta-drift, large-scale waves, spiral inflow, and general asymmetries associated with spiral bands.
2. Dominica and its instrumentation
Dominica is part of the volcanic chain of the Lesser Antilles. The island terrain forms a relatively simple north–south ridge but with higher peaks in the north and south reaching to about 1400 m. The dimensions of the island are about 17 km × 45 km. Under normal trade-wind conditions, the island receives about 7 m of precipitation per year on the highest ground and less than 2 m on the east and west coasts.
In March 2007, the exploratory phase of the Yale Dominica Precipitation Project began with the installation of four HOBO tipping-bucket rain gauges across the high mountains in the southern part of the island (Table 2; Fig. 3). These gauges record the time of each tip corresponding to 0.2 mm of precipitation. When data was downloaded on 29 August 2007, the logger clocks were accurate to better than one minute, allowing accurate analysis of rain rate during Dean. The location of the Freshwater Lake station is taken as a reference point in later diagrams.
The accumulated precipitation from the four gauges for 17 August is shown in Fig. 4. Except for a few brief pulses from 0200 to 1000 UTC, the heavy rain did not begin until 1100 UTC. During the full event, the two elevated stations (Freshwater Lake and Springfield) received about twice the amount at the coasts (La Plaine and Canefield).
One of the aspects of Dominica that makes it an effective natural laboratory is its proximity to the Météo-France weather radars on Guadeloupe and Martinique (Fig. 1). The Guadeloupe radar is especially helpful, as it has very little beam blockage to the south. The 5-min, 200-km Guadeloupe radar sweep is used in this paper with the relationship
from Regina (2007), where Z is the radar reflectivity and R is the rain rate (mm h−1). The coefficients in (2) were optimized for rainfall in the Lesser Antilles. For the standard 0.5° plan position indicator (PPI) sweep, the beam is about 1.5 km MSL at the distance of Dominica. A moderate ground echo can be detected from the northern peaks (Morne Diablotins, 1447 m) and occasionally from the southern peaks (Morne Trois Pitons, 1424 m). Echo from the southern peaks was removed in our analysis. The Guadeloupe radar is not very sensitive to low precipitation rates (the reflection threshold is 28 dBZ, or 5 mm h−1 with our Z–R relation), but this was not a serious drawback in the present application. The Martinique radar has more beam blockage than the Guadeloupe radar, but it is more sensitive and was closer to Dean’s eye.
A summary of the orographic effect on precipitation is shown in Fig. 5, an east–west distance–time (i.e., Hovmöller) diagram based on the Guadeloupe radar. During the morning, the region was dry except for a few traveling cells and an outer convective band that passed Dominica at 0530 UTC. In the afternoon, the precipitation was much heavier. The precipitating cells amplified over the island, but their motion is not seen in Fig. 5 after 1800 UTC because their south-southeasterly drift is not in the plane of the Hovmöller diagram.
The three soundings from Guadeloupe are also informative (Table 3). At 0000 UTC, the sounding resembled a summer fair-weather sounding with a trade-wind inversion, dry midlevels, and moderately high values of CAPE. The east-northeast winds of 22 m s−1 were more than double the usual trade-wind speed. At 1200 UTC, the winds were due easterly and nearly 35 m s−1. The air was moist at all levels. By 2400 UTC, the winds were weaker and from the southeast. The profile was moist but with reduced CAPE.
On the basis of Figs. 4 and 5, we identify three phases of the hurricane passage: (i) the quiescent period before 0400 UTC, with a few brief pulses; (ii) the narrow outer convective band crossing the island at 0545 UTC; and (iii) the long rainy period after 1100 UTC. Embedded within the rainy period is the intense period from 1400 to 1700 UTC, during which a convective band remained aligned with the island.
3. Three phases of precipitation
a. Phase 1: The initial quiescent period (0000–0400 UTC)
This initial quiescent period is summarized by an accumulated rain map (Fig. 6), produced by adding all the radar images in the 4-h period. It shows several cells forming more than 80-km upstream and advecting toward the island. A complementary presentation is given in Fig. 7, a Hovmöller diagram aligned with the cell drift direction (60°) that shows two cells passing over Freshwater Lake. One minor cell formed closer to the island but brought little precipitation. The rain intensity from each cell increased as the cells hit the island. From Fig. 6, we estimate the orographic enhancement factor for this phase to be in the range from three to six. Note that the rain gauge totals for this phase in Table 2 do not reflect this enhancement as a result of the rain gauges’ nonalignment with the drift direction.
b. Phase 2: The outer band (0530–0545 UTC)
The passage of the outer convective band of Hurricane Dean is shown in Fig. 8. By adding radar images taken at ∼1-h intervals, we see the band’s motion and continuity. The band was so narrow and fast moving that the 5-min radar scans give a poor time series of precipitation rate at any one location. The rain gauge time series show the band’s passage nicely, however (Fig. 9). The east coast station (La Plaine) registers the band first, starting at 0533 UTC and stopping after 3 min. Subsequently, the other stations measure brief pulses, ending with Canefield on the west coast. The band took 14 min to cross the island. Note that the two elevated stations received twice as much rain as the coastal stations (Table 2).
c. Phase 3: The rainy period (1100–2400 UTC)
Starting at 1100 UTC, heavily precipitating spiral bands near Dean’s center moved over Dominica. Over a 13-h period, the rain continued while the wind slowly completed its turn from southeast to south-southeast. The heaviest rain from 1400 to 1700 UTC was due to a band that remained nearly stationary over the island. Later, from 1700 to 2400 UTC, this band remained over the island and continued precipitating at a lower rate. The next section describes both the structure of this lingering band and the local orographic enhancement of the precipitation.
4. The stationary rainband in Hurricane Dean
The occurrence and structure of rainbands in Atlantic tropical cyclones have been discussed by Willoughby et al. (1982, 1984), Barnes et al. (1983), Jorgensen (1984), and Hence and Houze (2008). According to these reports, rainbands are spiral lines of flow convergence and lifting, with the updrafts tilting outwards from the storm center. They are more convective in the inflow regions and more stratiform downwind. They are approximately aligned with the spiral inflow in the hurricane boundary layer.
The banded structure of Dean is shown in Fig. 10, comprising the 200-km PPI from Martinique at 1445 UTC and the corresponding thermal IR MODIS. The radar image is easier to interpret (Fig. 10a). Near the western edge, the eye position C is evident at a point near to the NHC estimate at 1500 UTC (Table 1). The two eyewall clouds are neither axisymmetric nor complete. The primary spiral rainband positions A–B is located about 200 km east of the eye. Its length is about 300 km. At its south end, it is composed of three parallel rows of convection with a spacing of about 20 km. North of the Martinique radar, the band is slightly simpler and smoother in its structure. To a first approximation, convection elements move along the rainband from south to north.
The rainband is also seen in the MODIS thermal IR (TIR) image (Fig. 10b) but less clearly. A small circular structure (C) near the western edge of the MODIS cutout may be associated with Dean’s eye, but it is shifted 20 km to the south. Also, this ring is relatively warm in its annulus and cold in its center, the opposite of a cloud-free eye structure. The spiral band is also seen as a north–south line (A–B), with two big and five small, discrete circular overshooting convective tops. The largest convective top (near A) has a diameter of about 10 km.
In the 12-h period following this image (i.e., the rest of phase 3), the band changed its structure and orientation but roughly maintained its position in earth coordinates. It neither moved westward with the cyclone center nor northward with the cyclone winds. It remained aimed accurately at Dominica. This behavior is seen in the rainfall accumulation maps for 1100–1400, 1400–1700, and 1700–2100 UTC from the Guadeloupe radar (Fig. 11). In all three phases, the upwind end of the rainband precipitation was a point about 80 km southeast of the island. We find it unlikely that the island had any influence on the position of the rainband. This remarkable but “accidental” stationarity is partly responsible for the large total rainfall accumulations on Dominica.
5. Orographic enhancement of precipitation
The local orographic enhancement during phase 3 can be seen in the three accumulated precipitation maps (Fig. 11). All three examples show an approximate doubling of the precipitation over the island, especially over the southern hills. We show three long-interval maps to reduce the likelihood that precipitation maximum over the southern hills was an accidental result of a strong migrating precipitating cell. Note also the leeside rain shadow over the sea in the images in Fig. 11. The rain shadow is mostly west of the island in Fig. 11a, while in Figs. 11b and 11c it is located north and northwest of the island. This shift is consistent with the rotation of the wind during this phase. The rain shadow is almost completely free of precipitation.
The orographic enhancement or precipitation can also be seen in an east–west cross section for the period 1100–2400 UTC (Fig. 12). The radar transect across the southern mountains (with our Z–R relation) shows good agreement with the gauges. At 30 km to the east and west of the island, the accumulated precipitation was only 50 mm, while the mountain crest received about 500 mm. This large enhancement was the combined effect of the stationarity of the rainband and the orographic enhancement. To separate these two factors, we plot the precipitation along another east–west transect located 15 km south of the island, crossing the inflow and rainband (Fig. 12). It shows a broad peak with 200 mm of precipitation. From this comparison, it seems that the rainband stationarity and the orographic enhancement were both important in creating the heavy Dominica rainfall.
In the absence of orographic triggering of convection, it seems likely that some type of island-scale precipitation enhancement was working over Dominica. Given the short time response required, the seeder–feeder mechanism proposed by Bergeron (1968; see also Smith 1979) is a likely candidate. In this mechanism, falling raindrops gather terrain-induced cloud water as they fall to earth, increasing the precipitation rate over high ground. Quantitative estimates of this effect are given by Storebø (1976), Bader and Roach (1977), and Carruthers and Choularton (1983). A fundamental problem with these seeder–feeder models is that they do not predict a rainfall minimum in the lee (see Fig. 11). To capture this “dipole” feature in precipitation, we must use a different conceptual and mathematical model.
We simulate the dipole effect using a linear model of orographic precipitation (Smith and Barstad 2004). This model captures the influence of ascent and descent on cloud water density. It also advects liquid water downwind during conversion and fallout. Several assumptions of the linear model are met in phase 3 of Dean. First, the nonlinearity parameter is small (i.e., Nh/U < 0.1) if a small positive moist stability (N), a typical mountain height (h), and wind speed (U) are used. Second, the airstream remains nearly saturated over and around the ridge. Third, there is a background precipitation that is modulated by the terrain effect. While designed with stratiform clouds in mind (e.g., Crochet et al. 2007), the linear model treats the advection of any cloud system in which precipitation is formed by a generic two-step process. To apply it to the feeder–seeder scenario, we interpret these steps as the accretion growth of raindrops and their fallout. These processes are parameterized with two cloud delay times (τC and τF). The model cannot capture the inhomogeneous background precipitation field associated with the rainband.
The linear model uses simple stratified nonhydrostatic dynamics to lift the airflow and convert excess water vapor to cloud droplets. For a precipitation dipole to occur, the conversion and fallout must happen faster than the advection time over the ridge—that is, both cloud delay times must satisfy τ ≤ a/U, where a is the ridge width and U is the wind speed. For Dean, the parcel advection time over Dominica is about 30 000 m (30 m s−1)−1 = 1000 s, so a τ < 1000 s is suggested.
For the model run shown in Fig. 13, we chose a simple terrain consisting of two 1-km high Gaussian peaks, overlapped to form a ridge of Dominica’s dimensions. The wind is blowing from the south-southeast [U = 30 m s−1, wind direction (WD) = 155°]. A background precipitation of 20 mm h−1 is assumed (Fig. 5). Background parameters are air density (ρ = 1 kg m−3), specific humidity (qV = 0.017), and moist layer depth (HW = 3000 m), giving a water vapor flux of ρqVUH = 1530 kg m−1 s−1, similar to that in the 1200 UTC sounding (Table 3). Choosing a cloud delay time of τ = τC = τF = 400 s gives a good qualitative agreement with the maximum–minimum dipole seen in the radar maps in Fig. 11, with a precipitation maximum of 50 mm h−1 and a minimum of 2 mm h−1. The overall enhancement factor from the model is about 2.5. The relatively short cloud delay time (τF = 400 s) is consistent with a fast low-level scavenging process.
The primary difference between the linear model and the standard seeder–feeder model is in the cloud water budget. In the standard model, cloud water is created by uplift and reduced by scavenging and descent, but its concentration never goes negative. In the linear model on the other hand, there is assumed to be a preexisting cloud with cloud water and precipitation. Cloud water is increased by uplift and decreased by scavenging and descent, so that the perturbation cloud water concentration goes positive then negative (Fig. 14). Wherever the perturbation is positive, conversion to rainwater will be faster. Wherever the cloud water perturbation is negative, the conversion to rainwater will be slower. Thus, in the linear model, a single hill will produce a dipole of precipitation rate with respect to the background rate.
One further comment is useful. The leeside dry spot in the linear model is not caused by a net drying of the airstream, although this may be occurring. Rather, the dry spot is caused by the evaporation of cloud water as a result of leeside descent. This reduction in cloud water slows the accretion process and reduces the downwind rain rate.
6. Orographic enhancement and landslides
A preliminary report of damage to Dominica from Dean is given by FAO (2007). Because of the modest storm intensity and lack of proximity, the wind damage from Dean was not as intense as some hurricanes in recent history (Lugo et al. 1983; Tanner et al. 1991). With the exception of the banana crop, most of the damage on the island came from landslides in the steep volcanic terrain. Two deaths were reported and 150 homes were lost from landslides.
As landslides are caused by heavy rains persisting over time (DeGraff 1991; Larsen and Simon 1993; Iverson 2000), landslide prediction must take into account both intensity and duration of precipitation. Using data from volcanic terrain in Puerto Rico, Larsen and Simon proposed a threshold relationship between rainfall intensity and duration: I = 91.46D−0.86, where D and I are the duration (h) and the intensity (mm h−1). According to this formula, a rainfall intensity greater than 91.46 mm h−1 over one hour will trigger landslides. If the duration is 13 h, the critical rainfall intensity is 11 mm h−1 with a total amount of 143 mm. In Fig. 12 we see that with no orographic enhancement, the 13 h of accumulated precipitation would have been in the range of 100–200 mm—close to the landslide threshold. In reality, the actual precipitation on the high terrain exceeded 500 mm—several times the landslide threshold. Clearly, the orographic enhancement pushed the terrain past its landslide threshold. Evidently, the forecasting of landslides should not be based on open-sea precipitation estimates. Orographic enhancement of precipitation must be accounted for in forecasts of landslides.
Although Hurricane Dean was only a category-1–2 storm and it missed Dominica by 92 km, it still brought heavy rainfall accumulations to Dominica. This heavy rainfall had two causes: First, a nearly stationary spiral band remained focused on Dominica for nearly 12 h, even as the hurricane center was drifting away to the west. Second, a local precipitation enhancement mechanism was at work, roughly doubling the rainfall on the higher terrain.
While hurricane precipitation is mostly convective in origin, we find no evidence that the island triggered new convection in any phase of the hurricane event. This finding is in contrast to the normal climatology of Dominica, in which the triggering of convection over the windward east coast is responsible for most of the island’s precipitation. We attribute the lack of convective triggering in the hurricane to the high wind speed. First, using estimates of the gravity wave speed on the trade-wind inversion, we compute a Froude number (Fr = U/g′H) of 1.5. This means that information about the island could not be transmitted upstream by small amplitude waves, thus preventing upstream triggering. Second, the time for a parcel to move from the coast to the high terrain T = a/U = (8000 m)(30 m s−1)−1 = 266 s is not sufficient for the growth of a convective cell and the generation of its precipitation. Note that a parcel rising at 5 m s−1 would require 400 s to create a precipitating cell 2000 m high. Any cell failing to precipitate in this time will lose its chance as leeside descent begins.
The seeder–feeder mechanism, on the other hand, is much faster. As air rises over the terrain, the condensation of vapor to cloud droplet is virtually instantaneous. The accretion of cloud water onto existing falling raindrops is also quite quick; a drop falling through a 2-km layer of cloud with a fall speed of 6 m s−1 reaches the ground in T = (2000 m)(6 m s−1)−1 = 330 s. During fallout, the wind will carry the drops downwind a distance, d = 330(30) ≈ 10 km. When cloud delay times like this are used in the linear advection model, we obtain a realistic wet–dry dipole.
One further clue to the nature of the low-level ascent is the temperature data from the four rain gauge sites. Data from three days are shown in Fig. 15. On 17 August the normal diurnal cycle is missing completely. During the event, the 800-m Freshwater Lake site is only 3.5°C cooler than the sea level sites of La Plaine and Canefield. This difference nearly matches the moist adiabatic lapse rate (–3.6°C km−1) at this temperature, suggesting that the low-level air is cloudy. With cloudy low-level air, the seeder–feeder mechanism is possible.
Our conclusion about the feeder–seeder mechanism in hurricanes agrees with Misumi (1996) concerning terrain enhancement in a 1993 typhoon over the Ohsumi Peninsula in Japan. His observations are confused, however, by a strong secondary maximum in the lee of the hills, a feature we did not find. Geerts et al. (2000) also found evidence of low-level enhancement over the terrain of the Dominican Republic during Hurricane George in 1998. In addition, they found that the terrain triggered new convection within the eye. Roux et al. (2004) used single Doppler radar to describe the passage of Hurricane Dina near the island of La Réunion in the Indian Ocean. The 3-km-high terrain enhanced the rainfall and partly diverted the winds. No leeside rain shadow was identified. Numerical model studies of this case are underway (S. Jolivet 2008, personal communication).
Numerical simulation of landfalling typhoons over mountainous coastal China (Li et al. 2007) or Taiwan (Wu et al. 2002) have shown large precipitation increases too, but these cases are fundamentally different from ours, in that the terrain is large enough to modify the larger structure of the typhoon. Other studies of mountain effects on hurricanes are concerned more with the effect on the hurricane structure and track (e.g., Wei Jen Chang 1982; Bender et al. 1985).
Thanks to the people and government of the Commonwealth of Dominica for hosting this research, especially the Divisions of Agriculture and Forestry. Special thanks to Arlington James of the Forestry Division, who helped to find sites for the gauges and shared existing rain collector data. Thanks also to Nathaniel Isaac from the Met Office, Nancy Osler from the Archbold Tropical Research and Education Center and Ricky Brumant from the Agricultural Training School. Discussions with Anthony Didlake were useful. Sigrid R.-P. Smith assisted with the field work. Larry Bonneau assisted with the image analysis. The MODIS image was provided by NASA. Comments from the three reviewers were helpful. This research was supported by Météo-France and by a grant to Yale University from the National Science Foundation (Grant ATM-112354).
Corresponding author address: Ronald B. Smith, Department of Geology and Geophysics, Yale University, P.O. Box 208109, New Haven, CT 06520-8109. Email: firstname.lastname@example.org