a. Best-track data

Intensity and location data are taken from postanalysis “best track” files produced by the NHC (Jarvinen et al. 1984). The NHC best track combines all available tropical cyclone data to produce files that contain maximum sustained (1-min average) wind speed in knots, rounded to the nearest 5-kt interval (1 kt = 0.514 m s⁻¹), location, and minimum sea level pressure, smoothed to 6-h intervals. For the purposes of this work, “intensity” is classified as the best-track maximum sustained wind speed.

b. Environmental data

Environmental data are derived from archived model output (Claudette and Isabel) or from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Dora) written to diagnostic files designed for SHIPS training (DeMaria and Kaplan 1994, 1999; DeMaria et al. 2005). Prior to 2000, the SHIPS diagnostic file obtained its environmental parameters from 6-hourly NCEP–NCAR reanalysis data with a 2.5° grid spacing. For 2000 and afterward, Global Forecasting System (GFS) 0-h, 2.0° grid model analyses are used. The most significant difference in SHIPS parameters between the two data sources is a noticeable dry bias in the reanalysis low-level moisture fields compared to the model analyses.

Climatological and environmental data associated with each tropical cyclone are sampled at 6-h intervals at 0000, 0600, 1200, and 1800 UTC, respectively, and each environmental parameter is computed at the observation time and for “future” times at 6-h intervals out to 120 h. All environmental fields are modified such that the circulation signature of the tropical cyclone itself is removed to produce a better estimate of a tropical cyclone’s surrounding environment (DeMaria and Kaplan 1999). The circulation is removed from the wind field by subtracting out the tropical cyclone vortex and recomputing a new wind field, which is assumed to be representative of the environment in which the tropical cyclone is embedded. This research relies on the “perfect prog” assumption, which assumes that the environment simulated by the model accurately represents the true atmosphere. The accuracy of GFS tropical cyclone track forecasts has little impact on the resulting predictors, since only 0-h analyses are used in this postanalysis. For real-time forecasts, track errors are present, which can lead to errors in the environmental fields and eventual SHIPS intensity change forecasts.

Individual environmental parameters are based on spatial averages of the resulting environmental data fields between 200 and 800 km from a tropical cyclone’s center. (Divergence and vorticity use 0–1000-km radii.) The sea surface temperature (SST), in degrees Celsius, along a tropical cyclone’s best track is derived from Reynolds’s SST analyses, which have a 1° spatial resolution and represent values averaged over a 1-week period from buoy and satellite SST measurements (Reynolds and Smith 1994). A more detailed explanation of these parameters as well as the rationale for their use can be found in DeMaria and Kaplan (1994, 1999) and DeMaria et al. (2005).

c. Microwave imagery

Passive microwave data were collected from the constellation of Defense Meteorological Satellite Program (DMSP) satellites and the Tropical Rainfall Measuring Mission (TRMM) satellite. The relevant sensors include the Special Sensor Microwave Imager (SSM/I; Hollinger et al. 1987), aboard the DMSP satellites and the TRMM Microwave Imager (TMI; Kummerow et al. 1998). Both sensors measure upwelling microwave radiation emitted from the earth’s surface and atmosphere with radiances then converted into brightness temperature values using a form of Planck’s law. Raindrops emit (absorb) radiation over a large portion of the microwave spectrum. Areas of significant rain (>3 mm h⁻¹) appear warmer than the surrounding environment due to the greater emission from raindrops than from a low-emissivity ocean background at these frequencies. Precipitation ice and large raindrops scatter (in the Mie regime) upwelling radiation, especially at higher frequencies such as 85 GHz. Increased scattering occurs in areas of large raindrop and/or precipitation ice concentrations, resulting in these areas appearing colder than the surrounding environment in the 85-Ghz imagery.

Passive microwave channels observed by the SSM/I sensor include horizontally (H) and vertically (V) polarized 19-, 22- (V only), 37-, and 85-GHz frequencies. Pixel resolution is a function of frequency with 85 GHz being the highest at 16 km × 14 km and 19 GHz being the lowest at 70 km × 45 km. Data are oversampled by the satellite producing approximately 12 km × 12 km high-resolution pixels for 85-GHz data with the remaining three channels oversampled to low-resolution 24 km × 24 km pixels. TRMM’s lower altitude (400 versus 850 km) improves the spatial resolution to 7 km × 5 km at 85 GHz and to 30 km × 18 km at 19 GHz (oversampled to 5 km × 5 km and 10 km × 10 km, respectively). For both instruments, vertically and horizontally polarized 85-GHz channels are combined to form polarization corrected brightness temperatures (PCT85), which reduces the effect of the background (e.g., land or water) on the meteorological signal being studied (Spencer et al. 1989).

Several previous studies have shown that passive microwave brightness temperatures (T_b) and/or derived rainfall-rate products have a strong correlation with current and future intensities (Rao and MacArthur 1994; Rodgers et al. 1994; Rodgers and Pierce 1995; Cecil and Zipser 1999; Bankert and Tag 2002). In particular, the 19-Ghz T_b in the inner core of a tropical cyclone have been noted as an estimator of inner-core diabatic heating, a key factor associated with the intensification process. SHIPS-MI incorporates this information using two parameters: mean and maximum horizontally polarized 19-GHz brightness temperatures. The mean parameter is used as a proxy for the intensity of inner-core precipitation and latent heating, while the maximum T_b serves as a measure of the intensity of more localized heavy rainfall regions. (In the discussion, 85-Ghz imagery is often used as the primary indicator of convective activity). These parameters are calculated within a 100-km radius from the center of a tropical cyclone. For a satellite overpass to be used, the tropical cyclone center must be at least 100 km from the edge of the microwave swath and 100 km from land.

d. SHIPS-MI

The passive microwave version of SHIPS used in this overview has been improved over the version presented in Jones et al. (2006). The primary difference is that SHIPS-MI is now trained on a much larger microwave overpass sample extending back to 1988 and including 2004 data. New SHIPS environmental diagnostic files with 6-hourly (versus 12 hourly) resolution allowed for the creation of forecast models at 6-h intervals rather than the 12-h intervals used previously. The new SHIPS-MI also produces forecasts out to 5 days (120 h) versus 3 days (72 h) for the old model. SHIPS-MI predictors were adjusted to better correspond with those of the current SHIPS model, in preparation for operational testing.

For the Atlantic, changes included the removal of latitude as an independent predictor (it remains in a nonlinear term with shear) and the addition of 200-hPa divergence, pressure at the steering level (where the environmental flow matches the storm motion vector), number of days from the climatological peak in seasonal tropical cyclone activity, and the zonal component of storm motion. For the eastern Pacific, latitude as an individual predictor is retained despite its absence in the operational version of SHIPS, and its high correlation with SST. The high statistical significance of latitude in the eastern Pacific sample required its inclusion in the final regression, and resulted in a reduction in model error at longer forecast times. The significance of the latitude term suggests that highly correlated environmental predictors, particularly SST, may not be accurately represented in SHIPS.

Other changes in the eastern Pacific included the removal of the 200-hPa zonal wind velocity and the meridional component of storm motion. Additions included the number of days from the climatological peak, initial 200-hPa divergence, 200-hPa temperature, and 850-hPa relative vorticity. Changes in the microwave predictors included the substitution of mean 0–100-km 19-GHz brightness temperature in place of the mean 85-GHz brightness temperature and the removal of the 85-GHz symmetry term. A listing and brief description of each predictor included in SHIPS and SHIPS-MI (for both basins) are given in Table 1.

Despite the changes made to both the Atlantic and eastern Pacific versions of SHIPS-MI, their overall characteristics remain similar to those presented in Jones et al. (2006). SHIPS-MI still performs best for substantially intensifying or weakening tropical cyclones, with the greatest improvement over SHIPS observed for 12–36-h forecasts. Dependent sample mean absolute errors for both the old and new formulations of SHIPS-MI are similar (Fig. 1). SHIPS-MI retains a ∼5% improvement over SHIPS, which is statistically significant at a 95% confidence level. Additional documentation for SHIPS-MI and detailed comparisons against operational forms of SHIPS are given in Jones et al. (2006).

The same atmospheric characteristics are often described by multiple predictors. For example, SHIPS and SHIPS-MI each have three predictors that include shear in some form. The intensity change contribution from each shear-related predictor is combined to form the overall “shear” contribution category. To derive the total contribution for all quasi-independent characteristics, each model predictor is assigned to one of five categories (Table 2). The categories include climatology, SST-Potential, vertical wind shear, other environmental conditions, and microwave. This breakdown allows for the determination of which environmental and inner-core conditions were driving the intensity forecasts, and whether or not those conditions were a true reflection of what was controlling the intensity change for each tropical cyclone. The contributions from each category are computed by summing the individual intensity change forecasts for each predictor in a particular category. For example, the vertical wind shear contribution would be the sum of the (SHRD, SHRDLAT, and MSWSHRD) forecast intensity change values. The “importance” of a particular category is then defined by calculating the percentage of the total intensity change forecast accounted for by that category. For example, if the two SST-related terms yield a net +10 kt to a particular forecast, the two microwave-based terms yield a net −10 kt to the forecast, and all others add 0 kt, then the SST and microwave categories each have a 50% contribution to the total forecast.

The three cases discussed here are included in the developmental samples for both SHIPS and SHIPS-MI. Removing a single storm from the developmental regression has been tested, and has a negligible effect on the coefficients out to 72 h. The inclusion of the case study storms in the developmental sample does not affect the results shown. Comparisons with SHIPS are based on the 2005 version of that model (which also includes the case study storms in its developmental sample), not the 1999 or 2003 real-time forecasts. Since the forecast sample size for each case study is small, statistical significance cannot be assigned to any of the differences between the various model forecasts. Forecasts errors derived during real-time testing of SHIPS-MI, which is totally independent of the developmental sample and uses forecast tracks and environmental fields instead of analyses, were generally 10%–15% greater than dependent sample errors.

The versions of SHIPS and SHIPS-MI referred to throughout this paper do not take into account the effect of land interaction on tropical cyclone intensity. An exponential decay function has been developed for storms that cross over land, which in real-time operation could be applied equally to SHIPS and SHIPS-MI (Kaplan and DeMaria 1995). Since an analysis of the decay function skill is beyond the scope of this work, all forecasts that occur while a tropical cyclone crosses land are removed from consideration.

a. Storm overview

Hurricane Claudette developed from an easterly tropical wave that exited the African coast on 1 July and traveled westward across the North Atlantic, developing persistent deep convection by 6 July (Lawrence et al. 2005). Satellite intensity estimates suggested a tropical depression had formed by 0000 UTC 7 July, but no closed circulation could be found by reconnaissance. As such, the disturbance remained classified as a tropical wave. Environmental conditions consisted of light deep-layer shear (∼10 kt) and warm SSTs (>27°C), which were favorable for intensification. A low-level circulation was found by reconnaissance aircraft near 1800 UTC 8 July. Since tropical storm force winds were already present, the NHC upgraded the wave to Tropical Storm Claudette at 1800 UTC 8 July (Fig. 2a). This and future discussions of environmental conditions are based on values reported in the SHIPS diagnostic file and notes present in the NHC storm report (Lawrence et al. 2005).

Claudette continued to intensify, reaching hurricane intensity around 1200 UTC 10 July. It weakened rapidly thereafter, losing much of its organization. Southwesterly shear had increased with the approach of an upper-level low, displacing precipitation, and convection down shear of the center. Claudette slowed and began to take a northwestward course into the Gulf of Mexico (Fig. 3). The approximate center crossed the northeastern tip of the Yucatan Peninsula late on 11 July as a weak tropical storm. Claudette emerged into the Gulf of Mexico on 12 July, slowing down and stalling by 13 July. Claudette began to slowly reorganize in the Gulf of Mexico with SSTs near 29°C, though shear remained present. Reintensification continued with Claudette regaining hurricane status at 0600 UTC 15 July, just as shear began to lessen. Movement became westward and Claudette continued to intensify until landfall along the Texas coast near Matagorda Island at 1530 UTC 15 July. At landfall, Claudette had intensified to 80 kt with a minimum sea level pressure of 979 hPa, its peak intensity.

b. Forecast analysis

SHIPS-MI forecasts are computed for 6-h intervals out to 120 h (Fig. 4a) with all forecasts initiated from the time of a microwave sensor overpass. SHIPS model forecasts are created using the 2005 season operational coefficients and do not represent those produced in real time during 2003. Where available, SHIPS forecasts do have infrared brightness temperature and oceanic heat content adjustments applied; otherwise, the adjustment is set to zero (DeMaria et al. 2005). Forecasts originating prior to 8 July when Claudette was still classified as a wave are not considered. Similarly, forecasts occurring while Claudette was over land (Yucatan and Texas) are not considered as the effects of land on intensity change are beyond the scope of this study.

Mean absolute errors for both SHIPS and SHIPS-MI range from near 5 kt for 12-h forecasts up to 20 kt for 72-h forecasts (Fig. 5a). SHIPS-MI has slightly lower errors out to 24 h, after which SHIPS is the superior model. Forecasts generated just after the low-level circulation was detected, near 1800 UTC 8 July, called for intensification with SHIPS-MI forecasting more rapid intensification as a result of above normal near-center precipitation. While the initial intensification forecast was correct, intensification halted abruptly on 9 July, leading to large overforecasting errors (Fig. 4a). SHIPS-MI forecasts generated on 9–10 July indicated gradual intensification, but failed to capture the sudden intensification to 70 kt on 10 July. Precipitation remained near the center, and as a result, SHIPS-MI forecasts were higher than SHIPS, but by only a few knots. Neither SHIPS nor SHIPS-MI correctly forecast the substantial weakening that occurred after 1200 UTC 10 July, 24 h prior to landfall on the Yucatan Peninsula.

Claudette failed to intensify after emerging into the Gulf of Mexico despite the presence of 29°C SSTs. However, substantial shear coupled with the lingering effects of land remained present. These factors led to a decrease in precipitation and convection such that, by 12 July, little precipitation existed within 100 km of the best-track center (Fig. 6). Given these conditions, both SHIPS and SHIPS-MI forecast very little intensification prior to landfall, leading to substantial underforecasting errors during this period. SHIPS and SHIPS-MI forecasts were generally within 5 kt of each other.

To extract the physical conditions that are the primary contributors to SHIPS-MI forecasts, the individual contributions to forecast intensity change from each of the predictors are combined into the quasi-independent categories defined in Table 2. Combining all Claudette forecasts, the microwave (precipitation) contribution is greatest between 12 and 36 h, accounting for nearly 30% of the total forecast (Fig. 7a). SST-Potential becomes the dominant contributor after 36 h and grows to account for nearly 40% of the total forecast after 60 h. Shear and other environmental predictors each account for between 10% and 20% of the total forecast at all forecast times, while climatological predictors account for the remainder of the forecast.

Figure 8a shows prior 24-h forecast intensity change as a function of date, broken down by predictor category and compared against best-track intensity change. The forecast initialization date and time are 24 h prior to the date and time given for each intensity change value. (The positive contribution from the model mean intensity change, ∼5 kt, is not shown.) The negative contribution from shear for forecasts verifying between 9 and 11 July is clearly evident. The remaining SST-Potential, microwave, environmental, and climatological predictors generally have small positive contributions. The microwave contribution was initially positive. The positive SST-Potential contribution (∼3 kt) remains nearly constant with time as a result of potential values remaining the same through a combination of differing initial intensities and SSTs. Above normal environmental 200-hPa divergence and 850-hPa relative vorticity are the primary factors in producing the initially positive environmental contribution. However, none of the predictor contributions became sufficiently positive to capture Claudette’s intensification into a hurricane on 10 July. Also, only continued shear foretold the prelandfall weakening observed.

After moving into the Gulf of Mexico, very little (<5 kt) 24-h intensity change was forecast. The small positive contribution from SST-Potential was offset by small negative contributions from microwave and other environmental predictors. The microwave contribution only becomes positive again on 14 July (for forecasts verifying on 15 July) as Claudette is intensifying and the inner-core precipitation is becoming more persistent. However, the overall increase in forecast intensity change was only a few knots. As during the previous intensification phase, SHIPS-MI was not able to adequately capture Claudette’s intensification into a hurricane prior to landfall in Texas.

The conditions leading to Claudette’s intensification on 10 and 14–15 July are not currently taken into account by SHIPS-MI (or SHIPS). An analysis of available atmospheric parameters shows an increase in relative eddy flux convergence prior to each intensification phase (Fig. 9). Eddy flux convergence is the name given to the process whereby the interaction of a tropical cyclone and an upper-tropospheric trough can impart an inward flux of cyclonic angular momentum, increasing the tropical cyclone intensity (Holland and Merrill 1984; DeMaria et al. 1993). The most favorable orientation for eddy flux convergence occurs when the trough is located several hundred kilometers to the north and west of the tropical cyclone center. An increase in eddy flux convergence is often associated with increasing shear, leaving only a small window of opportunity for tropical cyclone intensity to increase before the broader effects of shear take over. Claudette’s small circulation (tropical storm and hurricane force winds extending out only 60 and 10 km on 10 July) made it more responsive to changes in its surrounding environment, especially shear (DeMaria 1996; Wong and Chan 2004). Thus, a rapid increase in eddy flux convergence may have aided the intensification on 8, 10, and 14–15 July, with rapid weakening following 10 July once shear became dominant. Eddy flux convergence is currently not taken into account by SHIPS (though the original 1994 version did include it as a predictor). If it were to be included again, it would have likely improved short-term Claudette forecasts with its positive contribution to intensity. The addition of inner-core precipitation from microwave brightness temperatures increased the forecast accuracy slightly for short-term forecasts, but the microwave version of SHIPS was unable to forecast the intensity peaks observed on 10 July. The relative lack of inner-core precipitation after 10 July up to 14 July led to relatively small contributions from the microwave predictors, which did not turn substantially positive until after the final intensification was under way.

a. Storm overview

Isabel formed from a tropical wave that exited the African coast on 1 September moving westward (Lawrence et al. 2005). Satellite estimates indicated a tropical depression had formed by 0000 UTC 6 September, which was upgraded to Tropical Storm Isabel 6 h later (Fig. 10). Isabel began moving west-northwestward on 7 September, during which time it was undergoing a period of rapid intensification while surrounded by favorable oceanic and atmospheric conditions. Isabel reached an initial peak intensity of 115 kt, making it a minimal category 4 hurricane, on 9 September before weakening slightly thereafter (Fig. 2b). Isabel turned westward on 10 September and began to intensify again, reaching category 5 status on 11 September. Peak intensity occurred at 1800 UTC 11 September, with maximum winds estimated at 145 kt. For the next 2 days, Isabel’s intensity fluctuated between 130 and 140 kt while taking on the appearance of an “annular type” tropical cyclone.

Isabel headed northwest on 15 September and began to weaken substantially, falling below major hurricane status by 16 September as shear began to increase. Isabel remained on a northwest heading, maintaining category 2 status until making landfall at Drum Inlet, North Carolina, at 1700 UTC 18 September with an intensity at landfall of 90 kt. Once inland, Isabel weakened and lost its tropical characteristics, becoming an extratropical cyclone in Canada by 20 September.

b. Forecast analysis

Both SHIPS and SHIPS-MI generally underforecast intensity during both intensification phases and afterward when Isabel maintained category 4 status (Fig. 4b). Only when Isabel weakened substantially after 15 September did the models overforecast intensity, though they correctly forecast substantial weakening. When all Isabel forecasts are averaged, SHIPS outperformed SHIPS-MI out to 60 h (Fig. 5b). The mean absolute error at this time for both models slightly exceeds 15 kt. The greatest differences between SHIPS and SHIPS-MI are apparent for forecasts initialized prior to 7 September. (Note that all forecasts are not shown.) Here, SHIPS-MI generally forecast a greater rate of intensification than SHIPS as a result of the abundant inner-core precipitation observed with the passive microwave data.

Neither model accurately forecast the second intensification period on 10 September. After Isabel reached its peak intensity, both models forecast substantial weakening, which failed to verify. SHIPS-MI forecast slightly more weakening than SHIPS, resulting in its slightly higher forecast error. The weaker SHIPS-MI forecasts during this period were not a result of a negative contribution from microwave predictors, but rather they were a result of the complex interaction between the highly correlated intensity, potential, and microwave predictors. During this period (11–15 September), Isabel transitioned into an annular-type hurricane. An annular hurricane is one in which outer rainbands dissipate, leaving a large, symmetrical annulus of precipitation surrounding a well-formed eyewall (Knaff et al. 2003). Annular hurricanes often maintain their intensity for long periods of time, even in the presence of less than favorable environmental conditions. Small fluctuations in intensity (±10 kt) occurred during this period, and are likely attributable to eye and eyewall processes such as the evolution of mesovortices within the eye (Kossin and Schubert 2004). Neither the annular-type status of Isabel nor its eye and eyewall characteristics are taken into account by SHIPS-MI. However, the training samples for both SHIPS and SHIPS-MI only contain a few long-lasting category 4–5 storms. As a result, they have a built-in weakening trend for storms that reach this intensity in all but the most ideal environmental conditions. The SST-Potential term has a strong negative contribution when a storm is near its MPI. The weakening forecast only verified when Isabel encountered higher shear on 15 September, disrupting its inner core.

Inner-core microwave predictors were the largest contributor to intensity change forecasts, accounting for over 30% of the total forecast out to 36 h (Fig. 7b). The large microwave contribution was the result of above normal inner-core precipitation present around Isabel after 9 September. Climatology, primarily in the form of above normal initial intensity, also accounted for more than 20% of the forecast out to 36 h. (Note that above normal intensity produces a negative intensity change forecast contribution.) The next largest contribution was from SST-Potential, which accounted for 20% of the forecast out to 24 h, increasing to nearly 50% at 72 h. Isabel nearly reached its maximum potential intensity on 11 September, resulting in a large negative intensity change forecast (especially at longer forecast times). Shear and other environmental predictors accounted for the remainder of the forecast, each accounting for less than 15% at all forecast times.

The large positive contribution from the microwave predictors is evident in Fig. 8b. These predictors combined to forecast almost 20 kt of 24-h intensification between 13 and 16 September. An example of the heavy inner-core precipitation present is given by an F-15 SSM/I microwave overpass occurring at 1423 UTC 14 September 2003 (Fig. 11). At this time Isabel had an annular appearance with heavy precipitation symmetrically surrounding a well-formed, large eyewall. Note that the higher-resolution 85-GHz imagery reveals a banding structure outside the eyewall, although the 19-GHz and infrared imagery (not shown) present a more annular structure. The mean 0–100-km average horizontally polarized 19-GHz brightness temperature at this time was 267 K, one of the highest values observed in the Atlantic sample. It is likely that much of the latent heating associated with the heavy precipitation went into maintaining and expanding the circulation rather than increasing Isabel’s maximum sustained winds. The large positive contribution from the microwave predictors partially offsets the negative contributions from SST-Potential and climatology. The SST-Potential and climatology terms are negative due to the initial intensity being substantially above normal (∼140 versus 57 kt). Isabel approached to within 5 kt of its oceanic maximum potential intensity on 11–12 September, and the forecast intensity contribution from SST-Potential 24 h later was at its most negative, −15 kt.

The shear contribution was slightly negative on 9 September due to a small increase in shear. Low-level vorticity (Z850) also briefly decreased to below normal levels, resulting in a ∼10 kt weakening signal from the environmental term. However, vorticity quickly recovered and the negative signal from the environmental category vanished by 10 August. Shear also lessened and the shear contribution again was slightly positive until 13 September. The positive overall contribution from weak shear was small, due to the high (negative) contribution from the “intensity times shear” term (MSWSHRD), offsetting the positive influence of weak shear. As a result, the lessening of shear did not translate into forecasts of substantial intensification. The contribution from the remaining environmental predictors was generally small, except after 16 September when the theta-e predictor, which represents a form of atmospheric instability, went above normal.

Considering the rare nature of category 5 hurricanes and those that reach their maximum potential intensity, SHIPS and SHIPS-MI performed rather well with Isabel (Fig. 4b). While both models underforecast the rapid intensification and annular hurricane phases, the total errors for Isabel were near training sample means. The positive microwave contribution was able to improve several of the early forecasts from 7 September (Fig. 8b). Neither model accurately captured the magnitude of the second rapid intensification (Fig. 4b). One reason was that the positive contribution from the decreasing shear was too small (as a result of offsetting terms) to add a substantial positive contribution to the forecast (Fig. 8b). Thereafter, both SHIPS models generally forecast little intensity change out to 24 h with substantial weakening forecast beyond 36 h. This was a result of very low SST-Potential values while Isabel approached its maximum potential intensity. The large negative contributions from SST-Potential and climatology were not completely offset by positive contributions from shear and microwave predictors despite very favorable inner-core and shear conditions. Only warmer SSTs (>29°C) would have significantly increased the forecast intensity change, but no evidence exists that warmer water was actually present. The problem was more a function of both SHIPS and SHIPS-MI containing relatively few long-lasting category 4–5 storms in their training samples, so regressions almost always forecast weakening for these storm types no matter what the surrounding conditions. The different training samples used by SHIPS and SHIPS-MI, along with the interaction of the many intensity-related predictors contained within both models, were likely key factors in the differences in forecast intensity during the annular phase of Isabel’s lifespan.

a. Storm overview

Convective activity from a tropical wave in the eastern Pacific increased, leading to the formation of a tropical depression by 0000 UTC 6 August. It was centered 540 km south of Acapulco, Mexico, moving northwest at approximately 10 kt (Fig. 12; Beven and Franklin 2004). The surrounding environmental conditions consisted of warm SSTs (∼28°C), moderate northeasterly shear (∼15 kt), and substantial low-level moisture (RHLO > 70%). Shear weakened almost immediately and Dora steadily intensified into a hurricane by 1200 UTC 8 August (Fig. 2c). Dora continued to intensify, reaching a peak intensity of 120 kt with an estimated minimum sea level pressure of 943 hPa at 0000 UTC 12 August.

Dora underwent a period of substantial weakening, with intensity falling to 70 kt by 1200 UTC 14 August while crossing 140°W and into the central Pacific. A second intensification phase commenced on 15 August with Dora reaching a second peak intensity of 100 kt at 0600 UTC 16 August, at which time Dora was located approximately 370 km south of the big island of Hawaii. Because of Dora’s proximity to Hawaii, airborne reconnaissance was able to verify the secondary peak in intensity. Following this peak, Dora weakened below hurricane intensity by 18 August. Dora crossed the international date line (180°) on 20 August and moved into the western North Pacific as a weakening tropical storm. Weakening continued until the low-level circulation dissipated on 23 August.

b. Forecast analysis

Both SHIPS and SHIPS-MI generally underforecast Dora’s intensity, except during its weakened state on 14 August (Fig. 4c). (Forecasts after crossing into the western Pacific are not considered.) Mean absolute errors for SHIPS exceeded 20 kt by 36 h, and for SHIPS-MI approached 20 kt by 48 h (Fig. 5c). SHIPS-MI was the more accurate of the two models with lower mean absolute errors for all forecast times out to 72 h. Errors for SHIPS-MI were lower since SHIPS-MI generally forecast greater intensification (or less weakening) than did SHIPS. Both models correctly forecast short-term intensification prior to 10 August, though failed to continue intensification beyond 48 h. After Dora reached 115 kt on 11 August, both models forecasted substantial weakening, though Dora maintained that intensity instead, much like Isabel. Neither SHIPS nor SHIPS-MI handled the second intensification phase on 15 August, with SHIPS even forecasting additional short-term weakening.

The differences in the SHIPS and SHIPS-MI intensity change forecasts for Dora were primarily the result of two factors. First, the addition of inner-core microwave imagery provided a positive signal during the latter half of the initial intensification phase (Fig. 8c), when above normal inner-core precipitation was present between 9 and 13 August. The second is a result of the inclusion of the initial storm latitude as a climatological predictor in SHIPS-MI. Throughout most of Dora’s life, its initial latitude was generally below the sample mean (16.8°N). This imparted an additional intensification signal, on the order of several knots, to SHIPS-MI forecasts out to 19 August. (Note that the overall climatological contribution remains negative.)

Climatological predictors, which now include latitude, account for a larger portion of the intensity change forecast in the eastern Pacific than similar predictors do in the Atlantic. Shear and other environmental conditions were relatively constant in the eastern Pacific; thus, they do not have the impact on intensity change that they have in the Atlantic. As a result, initial intensity and persistence alone can produce reasonably accurate forecasts for many eastern Pacific tropical cyclones. For example, a weak tropical cyclone that is intensifying is very likely to continue to intensify. Similarly, a strong tropical cyclone that is not intensifying further is likely to weaken. The importance of these predictors was evident in Dora forecasts. They accounted for over 40% of the intensity change forecast out to 36 h (Fig. 7c). The microwave contribution followed in importance accounting for more than 20% of the forecast out to 36 h. The contribution from SST-Potential increased from only 10% at 12 h to near 20% after 48 h, while the shear and other environmental categories each accounted for less than 15% of the forecast.

During the initial intensification phase (6–8 August), the 24-h intensity change forecasts were driven by a positive climatological contribution including a positive previous 12-h intensity change (persistence) coupled with low initial intensity and latitude (Fig. 8c). These combined to forecast >10 kt of intensification. Above normal theta-e excess (EPOS) led to the initial positive contribution from the environmental predictor grouping. The microwave contribution was near zero or weakly negative as a result of inner-core precipitation being enhanced southwest of the low-level center due to ∼15 kt northeasterly (200–850 hPa) shear. The shear contribution itself was negligible, as the importance of the shear predictors relative to SST-Potential, climatology, and microwave predictors is small for 24-h forecasts in the eastern Pacific. The SST-Potential contribution was also small as SSTs were near normal values (∼28°C).

During the latter half of the initial intensification phase after 9 August, the inner-core precipitation rapidly increased, imparting a positive intensity change contribution (Fig. 8c). At the same time, the increasing initial intensity and decreasing theta-e excess lowered the positive contributions from the climatological and environmental groupings. As a result, SHIPS-MI forecast less 24-h intensification than it had previously, reducing the forecast’s accuracy. When intensification ceased on late 10 August, inner-core precipitation was well established, leading to a positive microwave contribution of ∼10 kt. However, Dora began to move over cooler (∼27°C) SSTs, leading to an increasingly negative contribution from SST-Potential. With persistence now negative and the initial intensity high, the climatological contribution also turned strongly negative. The result was that SHIPS-MI and SHIPS incorrectly forecasted 24-h weakening for forecasts initialized after 11 August, when a forecast of little intensity change was needed. Still, the microwave predictors in SHIPS-MI increased its forecast intensity relative to SHIPS by several knots.

The weakening forecast finally verified when Dora weakened from 120 kt at 0600 UTC 13 August to 70 kt by 1200 UTC 14 August (Fig. 4c). Between 11 and 13 August, Dora moved over increasingly cooler SSTs, reaching 26°C by 13 August. The negative contribution from SST-Potential gradually increased, as did that of climatology due to negative persistence (Fig. 8c). Dora reintensified to 95 kt by 1200 UTC 15 August, despite little apparent improvement in the oceanic and atmospheric conditions. Since conditions were little changed, SHIPS-MI failed to forecast any intensification. Throughout the weakening and reintensification periods, the SHIPS-MI forecasts intensities after 48 h and were greater than corresponding SHIPS forecasts. This was primarily a result of the added influence of initial latitude in SHIPS-MI.

To an extent, both SHIPS and SHIPS-MI accurately captured the early intensification between 6 and 10 August. Afterward, both forecast substantial weakening while Dora maintained its 115–120-kt intensity. The incorrect weakening forecasts were a result of negative contributions from SST-Potential and more so from climatology. At the end of Dora’s initial intensification, it began moving over slightly cooler waters, which likely had an effect in the cessation of the intensification. Passive microwave imagery indicated a general decrease in the areal coverage and intensity of the inner-core precipitation. The decrease in inner-core precipitation was likely in response to cooler SSTs. However, SSTs remained warm enough through 13 August to support a 130-kt MPI. In the eastern Pacific as in the Atlantic, tropical cyclones rarely maintain category 4–5 hurricane status for long periods of time. Thus, SHIPS-MI in the eastern Pacific has a similar weakening bias for strong tropical cyclones.

On 13 August, Dora moved from >27°C SSTs to ∼26°C SSTs, representing a drop in MPI of 10 kt. Evidence from higher-resolution SSTs maps derived from microwave imagers indicated that Dora may have moved over a small region of 24°–25°C water, which would not be supportive of a major hurricane (not shown). In addition, high-resolution shear maps derived from satellite-tracked cloud movements indicated that the shear affecting Dora increased around this time as well (not shown). Given these factors and the small size of Dora’s circulation, rapid weakening ensued. While it was correctly forecast, it should be noted that, in fact, all forecasts made between 11 and 13 August called for weakening. The reintensification of Dora on 14–15 August was not forecast. The same evidence that showed Dora moving into colder SSTs and higher shear conditions than being sampled by SHIPS on 13 August also indicated movement back into a region of more favorable conditions. Dora’s small circulation allowed it to rapidly take advantage of the improved conditions with precipitation again becoming symmetrical around the inner core (Fig. 13). The small size of the precipitation core was such that pixels representing regions of little to no precipitation within 100 km of the center were being averaged into the mean microwave predictor at this time. The microwave terms therefore contributed little to the forecast during this period. Since the improvement in the environment and inner-core conditions was not adequately represented in SHIPS-MI, it could not account for them in its intensity change forecasts.