1. Introduction
Significant improvement has been achieved in the prediction of tropical cyclone (TC) intensity within the last two decades (Cangialosi et al. 2020) due in large part to statistical models like the Statistical Hurricane Intensity Prediction Scheme (SHIPS; DeMaria and Kaplan 1994) and the Logistic Growth Equation Model (LGEM; DeMaria 2009), as well as dynamical models like Hurricane Weather Research and Forecast system (HWRF; Surgi 2008). However, the errors in 24-h intensity forecasts have remained constant for the last 10 years (Cangialosi et al. 2020; Huang et al. 2021). Short-term errors in TC intensity forecasts can be impacted by errors in initial intensity (Emanuel and Zhang 2016). Since TCs spend most of their lifetime over the open ocean away from surface observations, satellite data are critical to estimate TC intensity. Some methods for estimating current TC intensity include the Advanced Dvorak Technique (ADT; Olander and Velden 2019) and AI-enhanced Advanced Dvorak Technique (AiDT; Olander et al. 2021) as well as Satellite Consensus (SATCON; Velden and Herndon 2020). Other studies have also used satellite data to estimate TC intensity, including Ritchie et al. (2012, 2014), Fetanat et al. (2013) and Zhao et al. (2016). Many recent studies have applied newer methods of machine learning to this same task (e.g., Zhuo and Tan 2021; Chen et al. 2019; Wimmers et al. 2019; Lee et al. 2020).
The purpose of this study is to predict current and short-term future TC intensity by combining satellite imagery and scalar features, like those used by AI-RI (Griffin et al. 2022), in a convolutional neural network (CNN). CNNs are designed to learn from spatial patterns, making them ideal for analyzing the convective organization of TCs as depicted in satellite-based infrared (IR) imagery. In addition to IR imagery, this study will also incorporate passive microwave (MW) imagery. Many studies have noted patterns in MW imagery related to current and future TC intensity (Alvey et al. 2015; Fischer et al. 2018; Tao and Jiang 2015; Harnos and Nesbitt 2016; Jiang et al. 2019), making it a well-suited addition to the CNN. Two different models are developed here: Deep Multispectral Intensity of TCs estimator (D-MINT), which utilizes MW and IR imagery, and Deep IR Intensity of TCs estimator (D-PRINT) which uses IR. Both models have utility because while D-MINT has more informative input data, D-PRINT has a higher and more reliable update frequency.
Recently, many studies have used CNNs to predict current TC intensity, as CNNs are designed to recognize objects and patterns in multiple images. Chen et al. (2019) utilized a CNN with a single IR brightness temperature (BT) image, a water vapor BT image, and passive microwave rain rate image. Wimmers et al. (2019) used 37- and 89-GHz satellite images. Lee et al. (2020) developed a model using two IR images at 12.0 and 10.8 μm, a 6.7-μm water vapor image, and a 3.7-μm shortwave IR image as inputs. However, none of these studies predicted future TC intensity trends, nor did any of these studies include environmental predictors in addition to satellite imagery. Machine learning has been used to enhance short term TC forecasts (Xu et al. 2021), but this study focuses on using current operational TC intensity predictions and does not explicitly use satellite imagery. Another method which has been used to predict future TC intensity in Pan et al. (2019) is a recurrent neural network, which is designed to interpret temporal or sequential information and produce the next output in a series. For example, Pan et al. (2019) uses TC location and intensity data from four previous times to predict a future intensity.
This paper is organized as follows. The TC and D-MINT/D-PRINT input feature data used in this study are described in section 2. Section 3 describes the CNN model, and section 4 described the method for determining which inputs to the CNN, known as features, are selected. The findings and a discussion are presented in section 5. Finally, a summary and conclusions are presented in section 6.
2. Data
a. Tropical cyclone database
This study predicts the current and future intensity [defined as 1-min average maximum sustained 10-m winds (MSW)] for global TCs, consistent with the National Hurricane Center (NHC) and Joint Typhoon Warning Center (JTWC) operational practices and best tracks. Global TCs are divided into five basins. Intensity and location histories from 1994 to 2021 for North Atlantic TCs and eastern North Pacific (ENP) are obtained from NHC best tracks. The intensity and location histories for western North Pacific (WNP) TCs, defined as west of 180°, as well north Indian Ocean (NIO) and Southern Hemisphere TCs are obtained from JTWC.
Three satellite-based TC current intensity estimates provided from the Cooperative Institute for Meteorological Satellite Studies and used operationally will serve as the baseline comparison: ADT (https://tropic.ssec.wisc.edu/misc/adt/info.html), AiDT (https://tropic.ssec.wisc.edu/real-time/adt/AiDT/aidt.html), and SATCON (https://tropic.ssec.wisc.edu/real-time/satcon/). SATCON is an ensemble model using ADT and microwave-based methods.
Future intensity predictions at 6, 12, 18, and 24 h will be compared to official forecasts from NHC and JTWC, as well as other operational models. These models include the following: SHIPS, LGEM, decay SHIPS (DSHIPS; DeMaria et al. 2005), HWRF, Hurricanes in a Multiscale Ocean-coupled Nonhydrostatic (HMON; Mehra et al. 2018), and ICON, an equal-weighted consensus of LGEM, DSHIPS, and interpolated HWRF and HMON. These deterministic forecasts for North Atlantic and ENP TCs are available in the “a-deck” files at https://ftp.nhc.noaa.gov/atcf/archive/. The a-deck files for other basins were provided by JTWC.
In addition to the explicit intensity forecasts mentioned above, D-MINT and D-PRINT are also compared to probabilistic forecasts of rapid intensification (RI). We use five conventional thresholds: an increase of 20 kt over a 12-h period; and 25, 30, 35, and 40 kt over a 24-h period (1 kt ≈ 0.51 m s−1). Current operational RI forecast guidance in the North Atlantic and ENP includes SHIPS-Consensus (Kaplan et al. 2015) and Deterministic to Probabilistic Statistical Model (DTOPS; Onderlinde and DeMaria 2018; DeMaria et al. 2021). RI forecasts will also be compared to AI-RI (Griffin et al. 2022), a CNN for predicting TC RI. In the WNP, NIO, and Southern Hemisphere, D-MINT and D-PRINT will be compared to the Rapid Intensification Prediction Aid (RIPA; Knaff et al. 2020). Instances where a TC center is over land within the RI lead time or the 12 h prior to the forecast time are omitted. RI forecasts are only analyzed at the synoptic hours of 0000, 0600, 1200, and 1800 UTC.
Systems categorized as an open wave or invest (based on best track data) are also omitted.
b. Satellite imagery
The IR satellite imagery used as inputs to D-MINT and D-PRINT in this analysis consists of BTs centered on the TC obtained from GOES, Himawari, MTSAT, and Meteosat depending on TC date and location (Table 1 with acronyms defined). Even though the central IR window wavelengths differ among these satellites, this difference is minimal and therefore there is no attempt at normalization based on the IR wavelength. Since the spatial resolution of these IR imagers varies from 2 to 4 km at nadir, the satellite IR data for each TC in this study is remapped to 128 × 128 grid points with a spatial resolution of 0.058° (∼5 km) using nearest-neighbor interpolation to homogenize the input data for the CNN. These satellite data are not parallax corrected, similar to Griffin et al. (2022), as the parallax error in the location of the highest-resolution BTs from 2019 to 2020 TCs is similar to the precision of the TC center latitude and longitude, and so issues in patterns of BTs due to parallax with respect to the TC center are deemed to be negligible. In addition to the current IR image, IR images of the TC going back every 3 h are also analyzed to assess their impact on intensity prediction. Seven different models are tested: one model with a 0-h IR image input (hereafter known as IR00), one model with the −3 and 0 h (IR03), etc. up to a model with −18-, −15-, …, −3-, and 0-h IR image input (IR18).
List of geostationary satellites used in this study. All geostationary latitude ranges are from 70°N to 70°S; GOES is Geostationary Operational Environmental Satellite and MTSAT is Multi-Functional Transport Satellite.
The MW satellite imagery used in this analysis is from the Microwave Imagery from NRL TC (MINT) dataset. This dataset, described in Cossuth et al. (2013), includes brightness temperatures from the DMSP SSM/I and SSMIS, TRMM TMI, Aqua AMSR-E, GPM GMI, and GCOM-W1 AMSR-2 satellites (Table 2, with acronyms defined). The diverse frequency bands from these sensors centered on 85, 89, and 91 GHz are normalized to 89 GHz using the technique of Yang et al. (2014). Small variations in spectral response between various sensors’ 37-GHz bands do not require renormalization. MW data for each TC is remapped to 64 × 64 grid points with a spatial resolution of 0.058° resolution using nearest-neighbor interpolation. At least 65% of the image box must be covered by the satellite overpass to be included in this analysis. Both the MW and IR satellite data are normalized by subtracting the mean BT and dividing by the standard deviation.
List of microwave satellites and sensors used in this study. DMSP: Defense Meteorological Satellites Program; SSM/I: Special Sensor Microwave Imager; H: Horizontal polarization; SSMIS: Special Sensor Microwave Imager/Sounder; TRMM: Tropical Rainfall Measuring Mission; TMI: TRMM Microwave Imager; GPM: Global Precipitation Measurement; GMI: GPM Microwave Instrument; AMSR-E: Advanced Microwave Scanning Radiometer—Earth Observing System; GCOM-W1: Global Change Observation Mission-Water; AMSR-2: Advanced Microwave Scanning Radiometer 2.
c. Scalar features
The scalar features used in this analysis are available from the SHIPS developmental data provided by the Cooperative Institute for Research in the Atmosphere (CIRA) available at https://rammb.cira.colostate.edu/research/tropical_cyclones/ships/developmental_data.asp. Scalar features are linearly interpolated to the D-MINT/D-PRINT image times from the previous synoptic time. A list of all the scalar features used in this analysis, chosen for consistency with AI-RI (Griffin et al. 2022), are shown in Table 3. Table 3 also serves as a reference for subsequent abbreviations used in the text. These features describe the location of the TC, as well as the surrounding environment and oceanic characteristics. However, unlike AI-RI, features such as current intensity and pressure, as well as the difference between MPI and current intensity, are not used in order to avoid obvious data leakage into the models. HIST, which includes past analyzed intensity, is included as it describes the TC age but not its current intensity. These scalar features are normalized from −1 to 1 for each oceanic basin. Data for the scalar features is either evaluated at time t = 0 or averaged over the entire forecast lead time as noted in Table 3. As MW images are not consistently available on synoptic times, scalar feature values are calculated from the lsdiag file prior to satellite image time. So, an IR00 or MW image at 0210 UTC would use the 0000 UTC lsdiag file to calculate the scalar features for all forecast lead times. The values of the scalar features are interpolated to be valid during the forecast time. So, since a time average of COHC for a 12-h forecast from 0000 UTC would be the average of the 0000, 0600, and 1200 UTC COHC, for an MW image at 0210 UTC, values of COHC from the 0000 UTC lsdiag file to interpolated to be valid at 0210, 0810, and 1410 UTC and then averaged. The value of V000 used for a 0210 UTC MW image, which is not time averaged, is linearly interpolated between 0000 and 0600 UTC from the 0000 UTC lsdiag file.
List of all features considered for D-MINT and D-PRINT.
For the results presented in section 4, D-MINT and D-PRINT will use scalar features from real-time lsdiag files provided by NHC and JTWC for consistent comparison with other operational forecasts. These real-time lsdiag files are used instead of the SHIPS developmental data described above, as their usage will provide a homogeneous comparison between D-MINT/D-PRINT and real-time TC intensity forecasts. As HIST is not included in the real-time lsdiag files, it is calculated from the working Best Tracks which can be found at http://hurricanes.ral.ucar.edu/realtime/plots/.
3. D-MINT and D-PRINT
A demonstration of the D-MINT and D-PRINT CNN model can be seen in Fig. 1, and a schematic flowchart of the algorithm is shown in Fig. 2. The configuration in Figs. 1 and 2 uses five IR satellite images centered on an agency-identified TC location over the previous 12 h (IR12) as inputs. Though Fig. 1 displays each IR image as an individual 128 × 128 × 1 input for clarity, the actual IR image input into D-MINT and D-PRINT in Fig. 1 is 128 × 128 × 5. D-MINT and D-PRINT begin by passing normalized IR data through a convolution layer. This layer produces an output, or “feature,” map (Fig. 1a). Convolution, in the context of deep learning, can be defined using Eq. (4) in Lagerquist et al. (2019), though the standard meaning of image processing through filtering also applies here. In D-MINT and D-PRINT, each convolution layer involves a convolution filter of 4 × 4 grid points. This filter operates spatially on the combined input grids to encode spatial patterns at higher levels of abstraction with each layer. Each filter has a different set of weights, which are initialized randomly. In D-MINT and D-PRINT, the first convolution layer produces a feature map with dimensions of 128 × 128 × 16. After every convolution layer, a “rectified linear unit” activation (ReLU; Maas et al. 2013) is applied. Activation is an elementwise nonlinear function applied to a feature map. Without it, D-MINT and D-PRINT would only learn linear relationships. The convolution and activation process explained above is repeated for 7 layers, until the final IR feature map is 4 × 4 × 128. This feature map is then flattened into a 1D vector with 2048 elements. This vector length and the convolution layer dimensions do not change with a different number of IR images.
Architecture of the D-MINT and D-PRINT convolutional neural network (CNN). The inputs used in this analysis are (a) normalized infrared (IR) BT data, (b) satellite microwave (MW) data, and (c) scalar predictors. For the normalized input and feature maps produced by convolution and pooling layers, positive values are in red and negative values are in blue. Dropout (0.2) randomly sets 20% of the input units in the concatenated data to zero during training to mitigate against overfitting.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
Schematic flowchart of D-MINT. This image was created at http://alexlenail.me/NN-SVG/AlexNet.html.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
The procedure explained above is repeated for the MW satellite features for D-MINT (Fig. 1b). MW features have an input dimension of 64 × 64 × 2, and five convolution and activation layers are used until the feature map is flattened to a 1D vector of 2048 elements. Finally, the flattened IR feature map, flattened MW feature map (D-MINT only), and input scalar features are concatenated together (Fig. 1c). These encoded data are then sent through one dropout to randomly zeroes out 20% of the layer’s values. Dropout reduces overfitting by forcing the weights in a given layer to evolve more independently (Hinton et al. 2012). The data are then sent through two dense fully connected layers to transform encoded data into predictions. The first dense layer uses the default ReLU activation function, which zeros out negative values as a computationally simple use of nonlinear operation. While this model configuration has similar aspects to Wimmers et al. (2019) and Griffin et al. (2022), such as using the default activation and depth of the CNN, certain hyperparameters such as dropout rate and learning rate were recursively searched when developing D-MINT and D-PRINT.
TC intensity prediction is a regression problem (i.e., we are predicting the actual TC intensity). To not only predict TC intensity but also display the uncertainty in the intensity prediction, the final dense layer in D-MINT and D-PRINT is 15 features, which produces a normal distribution of intensity for a given TC for 15 different quantiles: 1st, 2nd, 5th, 10th, 20th, …, 90th, 95th, 98th, and 99th. Therefore, the output from D-MINT and D-PRINT is 15 different TC intensities based on these quantiles, which portrays the uncertainty in the prediction. An example of the intensity output as a histogram is seen in Fig. 3. Any D-MINT and D-PRINT single-value intensity used in upcoming analysis is the average of the 30th–70th quantile, as that was found to be more accurate than using the 50th percentile based on the validation dataset.
Example of TC intensity histogram from D-MINT. The D-MINT average is the average of the 30th–70th percentile. MSW is the 1-min maximum sustained 10-m winds.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
4. Methodology
a. Data processing
The TCs used in this analysis are divided into three datasets: training, validation, and testing. The training dataset consists of global TCs from 1994 to 2015. Satellite images of these TCs are randomly rotated 40° and/or flipped to aid in generalized training; however, no image is duplicated to expand the dataset as was done in Griffin et al. (2022). The validation dataset is global TCs from 2016 to 2017. This dataset is used to determine the optimal scalar predictors and number of IR images for D-MINT and D-PRINT. Therefore, the testing dataset, which is 2019–21 global TCs, remains independent of any model training or design. Features in the testing dataset are also normalized based on TCs from 1994 to 2017. TCs from 2018 are not utilized in any dataset as the scalar features are not available globally. The number of TC times, by basin, for each dataset is given in Table 4. A “TC time” is a current intensity estimate with IR00, so these numbers do vary based on lead time and IR histories needed.
Number of TC times by basin, where a “TC time” is a current intensity estimate with IR00, so these numbers do vary based on lead time and IR histories needed.
b. Feature selection
To identify which scalar features are optimal at predicting TC intensity, we begin by training model configurations with all MW and scalar features listed in Table 3 and the current IR image (IR00) for forecast lead times between 6 and 24 h. This was done to streamline the list of optimal scalar features when predicting intensity. After these different forecast hours are trained, the optimal scalar features are selected using “permutation importance” (Jergensen et al. 2020) applied to the validation dataset. In permutation importance, input data for a given feature are randomly shuffled; for example, the RSST value from the 200th TC image is used in place of that for the 15th TC image, while leaving the other features’ input data unchanged. This is meant to reveal the sensitivity of the permutated feature on the prediction. The permutated feature’s impact on overall model skill is measured by calculating the root-mean-square error (RMSE) between the average of the 30th–70th quantiles and the operational best track current or future intensity. A scalar feature is deemed optimal if the RMSE of the permutated version is higher than that of the nonpermutated version from an average of 1000 instances of bootstrap sampling with replacement.
Once the optimal scalar features are identified, D-MINT and D-PRINT are trained again using only these scalar features and the corresponding satellite imagery. Since training a CNN iterates from randomly initialized weights, each training run produces nonidentical results. So, we trained five CNNs and use the model with the lowest RMSE on the validation dataset. The RMSE for the seven model configurations using different histories of IR images for the validation dataset can be seen in Fig. 4. The RMSE can vary with the number of IR images, though this comparison was kept to a homogeneous collection. Based on Fig. 4, it was decided to use IR12 for the 0-h forecast, IR06 for the 6-h forecast, and IR00 only for forecast lead times longer than 6 h. This applies for both D-MINT and D-PRINT. The satellite IR features and scalar predictors used by D-MINT and D-PRINT for all forecast lead times can be seen in Table 5.
RMSE for the validation dataset (2016/17 global TCs) based on IR image histories and forecast lead time. D-MINT is depicted in blue, and D-PRINT is depicted in red. The triangles represent the mean RMSE over all five attempts, with the squares indicating the minimum RMSE. The x-axis labels indicate the length of the IR history; for example, “IR09” means that the model version uses IR imagery from 0, 3, 6, and 9 h before the 0-h estimate time. The range of the y axis varies by forecast lead time.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
Satellite IR features and scalar features used by D-MINT and D-PRINT.
c. Future intensity prediction
Unfortunately, MW images are not conveniently available on synoptic times, unlike many operational models used for predicting future TC intensity. So, to compare D-MINT/D-PRINT to these operational models, the operational model intensities are linearly interpolated to be valid at the D-MINT/D-PRINT forecast time. For example, a 12-h forecast from an MW image at 0210 UTC is valid at 1410 UTC. D-MINT and D-PRINT are compared to a forecast when they are within ±3 h of the synoptic time. D-MINT/D-PRINT forecasts from 0300 (0301) UTC are then compared to 0000 (0600) UTC forecasts. For the example above, 12-h forecast from a MW image at 0210 UTC would be compared to operational forecasts from 0000 UTC valid at 1410. A 12-h forecast from a MW image at 0310 UTC would be compared to the operational forecast from 0600 UTC linearly interpolated to 1510 UTC. This is done because always comparing D-MINT/D-PRINT to the next possible forecast would put D-MINT/D-PRINT at a disadvantage, but always comparing D-MINT/D-PRINT to the previous forecast would give them an unfair advantage. This analysis is conducted for all available MW images.
d. Comparison to RI forecasts
Unlike the comparison to operation models mentioned above, RI forecasts should only be analyzed at synoptic hours, as RI probabilities are not easily interpolated. When predicting RI, forecasts from D-MINT and D-PRINT are only analyzed if a MW image is available within 2 h prior to or 1 h after the synoptic time. If multiple forecasts are available, RI is forecasted using the D-MINT forecast closest to the synoptic time without going over, unless the only available forecast is after the synoptic time.
Predicting RI from a deterministic model can be as simple as assigning a “yes” when the forecasted increase in intensity exceeds the RI threshold, and “no” when it does not. One option for assessment of the model is the Peirce skill score (PSS). The PSS is equal to the probability of (correct) detection (POD) minus the probability of false detection (POFD), and therefore does not discourage forecasting rare events (Wilks 2006). A PSS of 1 represents a perfect RI forecast (Wilks 2006). In addition, false alarm ratio (FAR) is also calculated. A 2 × 2 contingency table, as well as equations for each metric, can be found in Tables 6 and 7, respectively. Probabilistic RI predictions are converted to a yes–no forecast by calculating the RI probability where the POFD equals the climatological probability of RI for each model (as was done in Kaplan et al. 2015). Thus, this critical RI probability varies by TC basin, RI threshold, and RI forecasts (Table 8).
Relationship between forecasts and observations of RI in a 2 × 2 contingency table.
Critical probability thresholds (%) according to model, basin and each of the five RI thresholds.
e. Shapley additive explanation (SHAP) values
The impact of an individual feature on the predicted TC intensity will be assessed using the technique of Shapley additive explanation (SHAP) (Shapley 1953; Lundberg and Lee 2017; Lundberg et al. 2018, 2020). SHAP is an “explanation model,” using a large combinatorial analysis to estimate the relative contribution of each input feature to the corresponding output of a predictive model. While it relies on linear approximations of nonlinear models, SHAP has the advantage of supplying simple and interpretable solutions to questions of model performance. For a regression problem, SHAP values indicate the amount an individual input feature increases or decreases the predicted value from the mean. SHAP values are calculated using the shap python library, (https://github.com/slundberg/shap). For more information, refer to Mangalathu et al. (2020).
5. Results and discussion
a. Current intensity estimation
The error for estimating real-time intensity for the independent 2019 to 2021 Global TCs can be seen in Fig. 5. In Fig. 5a, black and gray bars represent the RMSE for D-MINT and D-PRINT, respectively. In all basins, the error is lower for D-MINT than D-PRINT, indicating that MW imagery undoubtedly adds skill. Both D-MINT and D-PRINT have a lower error than the ADT (pink bars) in all basins but NIO. This is consistent with Chen et al. (2019), who first developed a CNN with lower error than the ADT. D-MINT and D-PRINT also have a lower RMSE than AiDT (red bars) for all basins except the NIO and Southern Hemisphere. Compared to SATCON (purple bars), D-MINT and D-PRINT have lower error than SATCON in three out of five and two out of five basins, respectively.
(a) RMSE for estimating 2019–21 TCs current intensity by basin. (b) RMSE and (c) Bias by basin for estimating 2019–21 TCs current intensity based on previous 12-h MSW change (DELV): weakening (DELV ≤ −10 kt), steady (10 < DELV < 10 kt), and intensifying (DELV ≥ 10 kt).
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
A comparison of the error between D-MINT/D-PRINT and other published methods for estimating current TC MSW is given in Table 9. The RMSE for D-MINT and D-PRINT is lower than Chen et al. (2019), Wimmers et al. (2019), Lee et al. (2020), and Zhuo and Tan (2021) although these are not homogeneous comparisons.
RMSEs of several intensity estimation techniques. The RMSE for D-MINT, D-PRINT, ADT, and SATCON are from a homogeneous sample, while the RMSE for the other methods are as published.
Both D-MINT and D-PRINT use IR image input from the previous 12 h, so it is important to probe how these product’s skills are affected by any major 12-h trend in the IR imagery. To this end, Figs. 5b and 5c partition the product error by the previous 12-h intensity change (DELV). DELV is broken into three categories: weakening (DELV ≤ −10 kt), steady (10 < DELV < 10 kt), and intensifying (DELV ≥ 10 kt). As seen in Fig. 5b, the RMSE for D-MINT and D-PRINT is the lowest for steady TCs. Therefore, one hypothesis for D-MINT and D-PRINT having a higher RMSE than AiDT and SATCON for NIO and Southern Hemisphere TCs is that a higher percentage of estimates in these basins are for weakening and intensifying TCs (54.4% and 45.4%, respectively) compared to the North Atlantic (25.6%) and eastern (31.3%) and western (40.0%) North Pacific. Also, the bias in Fig. 5c does not depict a noticeable lag in TC intensity, as the bias is negative for both weakening and intensifying TCs. If the number of IR images were causing a lag in TC intensity, we would expect a positive bias for weakening TCs due to the IR images backward in time being from a stronger TC. The bias for D-MINT and D-PRINT are similar across all basins. The lowest magnitude bias for both occurs in the ENP, where consequently the RMSE is also the lowest. Both D-MINT and D-PRINT also have the most negative bias for weakening in the NIO and Southern Hemisphere, but the cause is unknown.
The RMSE and bias also differ based on observed TC intensity, as seen in Fig. 6. Figure 6 depicts the 15-kt running average of RMSE (left) and bias (right) across 5-kt bins. The number of TC images in each bin can be seen in the light gray and right y-axis. Globally (Fig. 6a), all methods generally overestimate the weakest TCs (MSW ≤ 40 kt) while underestimating the strongest TCs (MSW ≥ 140 kt). SATCON, ADT and AiDT generally have a higher RMSE and bias for weaker TCs than D-MINT and D-PRINT. Therefore, another likely explanation for the lower SATCON and AiDT RMSE in NIO and the Southern Hemisphere is that fewer of these weaker TCs exist in the corresponding dataset. Only 20.7% and 21.9% of NIO and Southern Hemisphere TCs have a MSW ≤ 40 kt, respectively, which is lower than the North Atlantic (33.2%) and eastern (48.8%) and western (35.1%) North Pacific.
(left) RMSE and (right) bias for estimating 2019–21 current intensity (MSW). The number of TC images in each bin can be seen in the light gray and right y axis. The range on the y axis for each panel varies by basin.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
Since the output from D-MINT and D-PRINT is a distribution of possible TC intensities, relative uncertainty in the estimation of TC intensity can be characterized. As seen in Figs. 7a and 7b, the width of the histogram for these intensity estimations is smaller for D-MINT than for D-PRINT. Thus, D-MINT has less uncertainty in its prediction. The average uncertainty in the current TC intensity prediction also varies with TC intensity. The most confident estimations occur for the weakest TCs (Figs. 7a,b), where there is less than 7 kt between the 30th and 70th quantile of predicted intensity (Fig. 7b). The histogram width, and the range of the 30th–70th percentile, then sharply increases until about 65 kt before increasing at a lesser rate. The strongest TCs differ by about 14 kt from the 30th and 70th percentile of predicted intensity (Fig. 7b). However, this difference between the 30th and 70th percentile does not appear to be due to the overall distribution of TC intensities, as the average histogram width for a 40-kt TC is similar to that of a 90-kt TC (Figs. 7c,d). Many other factors can cause uncertainty in the current intensity estimate prediction. One such factor is the previous 12-h intensity change, with larger magnitude changes resulting in greater uncertainty (Figs. 7e,f). Environmental factors, such as high SST and moderate to high wind shear, can also increase the uncertainty in the current intensity estimate prediction (not shown).
(a),(b) Average width and (c),(d) standard deviation of the (left) current TC intensity histogram and (right) 30th–70th percentile when estimating 2019–21 current intensity. (e),(f) Histogram width and width of the 30th–70th percentile based on previous 12-h intensity change.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
b. Future intensity prediction
The errors for predicting future TC intensity for 2019–21 North Atlantic and ENP TCs are given in Fig. 8. In Figs. 8a and 8b, the forecast lead time is based on the time of the D-MINT/D-PRINT imagery, so for example, a 12-h forecast from an image at 0400 UTC is valid at 1600 UTC. Because other forecasts are only issued at synoptic times, their intensities are linearly interpolated here to match the D-MINT and D-PRINT forecast. D-MINT and D-PRINT are compared to a forecast when they are within ±3 h of the synoptic time. Thus, D-MINT/D-PRINT from 0300 (0301) UTC is compared to 0000 (0600) UTC forecasts. In the North Atlantic (Fig. 8a) and ENP (Fig. 8b), the error for D-MINT and D-PRINT is higher than the official forecast from NHC, as expected. The error for D-MINT and D-PRINT, though, varies within the range of error for LGEM, HWRF, and HMON. D-MINT and D-PRINT also have a higher error compared to ICON. However, adding D-MINT as a member in the ICON consensus results in a consistently lower error. Figures 8c and 8d show the errors for products available at the synoptic time. Adding D-PRINT to ICON results in a lower error than ICON alone here as well.
RMSE for estimating 2019–21 TCs future intensity (MSW) for North Atlantic and eastern North Pacific TCs by (a),(b) microwave image time and (c),(d) synoptic time.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
Similarly, Fig. 9 shows the comparative error for various TC intensity forecasts for 2019 to 2021 WNP, NIO, and Southern Hemisphere TCs. In the WNP (Fig. 9a) and NIO (Fig. 9b), D-MINT has a lower error than all other forecast models, including the official JTWC forecast (except for the NIO 6-h forecast). D-MINT is less skillful than JTWC in the Southern Hemisphere (Fig. 9c), but more skillful than LGEM and HWRF for 18- and 24-h forecasts. D-PRINT is less skillful the official JTWC forecast for these basins but has more skill in the WNP than HWRF and LGEM (except the 6-h forecast).
RMSE for estimating 2019–21 TCs future intensity for the (a) western North Pacific, (b) north Indian Ocean, and (c) Southern Hemisphere.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
To the best of our knowledge, the other forecast methods in Figs. 8 and 9 use a baseline current intensity, potentially the working best track from NHC, JTWC or BoM, either as a direct input or a background assumption when predicting future intensity. Thus, D-MINT and D-PRINT have the added advantage of being a more independent tool for real-time TC diagnostics. Likewise, using the working best track intensity tends to incorporate the same systematic biases found in the best track, whereas D-MINT and D-PRINT are less prone to this bias by incorporating intensity history rather than real-time intensity. This means that the skill of D-MINT and D-PRINT is likely to be understated relative to the other methods.
To further assess the bias when predicting future intensity, Fig. 10 presents the bias based on the 12-h change in intensity prior to the valid time. For example, in Fig. 10 the intensity change for the 6-h forecast is calculated from t − 6 to t + 6, where t is the time of the forecast, and the intensity change for the 24-h forecast lead time is calculated from t + 12 to t + 24. For weakening TCs, D-MINT and D-PRINT have a negative bias for the 6-h forecast lead time for all basins. Therefore, the predicted intensity is too low compared to the best track. This is in contrast to all other models, which have a neutral or positive bias. As forecast lead time increases, the D-MINT and D-PRINT bias for weakening TCs increases and becomes positive (with exception of the NIO). Yet, D-MINT and D-PRINT generally continue to have a lower bias compared to other models for weakening TCs. Overall, the bias for weakening TCs from D-MINT and D-PRINT is closest to zero over all forecast lead times, aside from HWRF in the North Atlantic (Fig. 10a), HWRF and HMON in the ENP (Fig. 10b) and LGEM in the NIO and Southern Hemisphere (Figs. 10d,e). For steady TCs, D-MINT and D-PRINT also generally have a lower magnitude bias compared to other models for the WNP, and except for HWRF and ICON in the North Atlantic and ENP.
Bias by basin for estimating 2019–21 TCs future intensity based on previous 12-h MSW change prior to valid time: weakening (12-h MSW change ≤ −10 kt), steady (10 kt < 12-h MSW change < 10 kt), and intensifying (12-h MSW change ≥ 10 kt). For example, the MSW change for the 6-h forecast lead is calculated from t − 6 to t + 6, where t is the time of the forecast, and the MSW change for the 24-h forecast lead is calculated from t + 12 to t + 24. The bias range on the y axis varies by basin.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
For intensifying TCs, all models have a negative bias, meaning that the predicted intensity is low compared to the best track. For the ENP (Fig. 10b), D-MINT and D-PRINT have the least negative biases compared to other forecasts, and therefore their intensity forecasts best represent the best track intensity. In the North Atlantic (Fig. 10a) and WNP (Fig. 10c) HWRF and HMON have the least negative biases compared to other forecasts, followed by D-MINT. D-MINT has the least negative bias in the NIO (Fig. 10d) over all forecast lead times, and only HWRF has a less negative bias than D-PRINT. In Southern Hemisphere TCs, both JTWC and HWRF have a less negative bias than D-MINT and D-PRINT.
c. Rapid intensification
The Peirce skill score (PSS) for deterministically predicting (i.e., a yes–no forecast) 12- and 24-h RI is in Fig. 11. Here the pink bars indicate the PSS for SHIPS, which had a higher PSS for all lead times and basins compared to LGEM and DSHIPS. AI-RI, SHIPS Consensus, and DTOPS are only available for the North Atlantic and ENP. As seen in Figs. 11a and 11b, these RI probabilistic models have the highest skill. This is unsurprising, as these probabilistic models are specifically trained to predict RI. Compared to D-MINT, D-PRINT is generally less skillful for all lead times and basins. Therefore, MW imagery does add skill at predicting RI. This is consistent with Fischer et al. (2018), which found convective features in 37- and 85-GHz brightness temperatures associated with RI.
Peirce skill score (PSS) for predicting RI in 2019–21 TCs. A higher PSS indicates a more skillful RI prediction. The maximum value for the PSS is 1.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
For deterministic predictions of RI, the most skillful model depends on RI threshold and basin. In the North Atlantic (Fig. 11a) and ENP (Fig. 11b), the skill of D-MINT and D-PRINT compared to other deterministic models decreases with increasing RI magnitude. D-MINT is the second most skillful at predicting 20 kt (12 h)−1 RI in the North Atlantic, and the least skillful at predicting 40 kt (24 h)−1 RI. For the ENP, D-MINT and D-PRINT are generally less skillful compared to the North Atlantic, even though other deterministic models are more skillful. This difference in D-MINT skill could be due to the importance of satellite imagery when predicting RI in these basins. Of the 10 predictors used in the SHIPS-Rapid Intensification Index, satellite predictors were third and fifth most important in the North Atlantic but only sixth and ninth most important in the ENP (Kaplan et al. 2015). Therefore, the satellite predictors have less weight in the ENP than the North Atlantic.
The forecast skill over the WNP, NIO, and Southern Hemisphere can be seen in Figs. 11c–e. Probabilistic RIPA is not included because it greatly reduced the number of RI forecasts for a homogeneous sample. In these basins, HWRF is generally the most skillful at predicting RI, though RIPA is the most skillful at deterministically predicting RI when available (not shown). HWRF generally has the highest POD of all deterministic models for these basins, and the second highest FAR. However, the rarer the event, the greater the contribution to the PSS for a correct “yes” forecast (Wilks 2006), so HWRF has the highest PSS despite the increased number of incorrect “yes” forecasts. Over all RI thresholds in these three basins, D-MINT is the next most skillful deterministic model, with the second highest POD and lowest FAR. This is especially evident in the WNP (Fig. 11c). One hypothesis for increased skill in these basins compared to other models is that satellite predictors have a greater importance for predicting RI in these basins, especially in the WNP (Knaff et al. 2018). However, the D-MINT and D-PRINT skill generally decreases with increasing intensification rates over the 24-h RI thresholds, though HWRF’s skill remains relatively constant. This is consistent with how satellite predictors lose importance in RIPA with increasing 24-h RI thresholds in favor of environmental factors.
A skill comparison of D-MINT and D-PRINT to other probabilistic RI forecasts can be seen in the BSS in Fig. 12. Unlike the PSS in Fig. 11, which assessed deterministic (yes–no) RI forecasts, D-MINT is competitive and often more skillful than the probabilistic models that are specifically trained to predict RI. This increased skill when using MW imagery is consistent with Rozoff et al. (2015). D-MINT is generally more skillful for all lead times and basins than D-PRINT.
The Brier skill score (BSS) compared to climatology for predicting RI in 2019–21 TCs. A higher BSS indicates a more skillful RI prediction. The maximum value for the BSS is 1.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
For TCs in the North Atlantic (Fig. 12a), D-MINT has the highest BSS for 20-, 25-, and 30-kt RI, and therefore is the most skillful. The BS for D-MINT is significantly higher than that of DTOPS to >90% on a one-tailed f test for 25-, 30-, and 40-kt RI, and that the sample size is too small for statistical significance in the other cases. Compared to the other models, D-MINT also has a higher probability of RI for at least half of the instances of RI occurring for all RI thresholds (not shown). However, the improvement in skill from this higher relative probability of RI for positive cases is heavily offset by the even higher relative probability of RI for negative cases compared to the other models. Then at 35 and 40 kt, D-MINT has less relative skill because the difference in RI probability is both lower for these thresholds when RI does occur, and higher when RI does not occur.
For the ENP (Fig. 12b), DTOPS is the most skillful for all RI thresholds, consistent with Griffin et al. (2022). Latitude is a high contributor for DTOPS (Onderlinde and DeMaria 2018). So, for TCs north of 15°N, D-MINT has the highest BSS of all models over 24-h RI (not shown). This increased skill is consistent with Griffin et al. (2022). Like AI-RI, D-MINT and D-PRINT also include latitude as a feature, but its relative impact is much lower than in DTOPS. The increased skill for D-MINT compared to DTOPS is not only due to increasing RI probability when RI does occur, but also decreasing RI probability when RI does not occur for given thresholds. D-MINT and D-PRINT are more skillful than SHIPS Consensus for 25- and 40-kt RI and more skillful than AI-RI for 40-kt RI.
Based on the BSS, D-MINT is more skillful than RIPA for predicting all 24-h RI thresholds in the WNP (Fig. 12c). The BSS for D-MINT is significantly higher than that of RIPA to >90% on a one-tailed f test for 25-, 30-, 35-, and 40-kt RI. On average, D-MINT also has a higher probability of RI compared to RIPA when 25- and 30-kt RI occurs, and a lower probability for all other instances. This average difference, though, is less than 6%. D-MINT does have a higher probability of RI than RIPA for over half of 20-, 25-, and 30-kt RI occurring. In NIO TCs (Fig. 12d), D-MINT is more skillful than RIPA for 20-kt RI, but is less skillful for 24-h RI. On average, D-MINT has a lower probability of RI than RIPA when 24-h RI occurs. Combined with the increased frequency of 24-h RI in the NIO compared to other basins, these lower probabilities result in D-MINT being less skillful. For 20-kt RI, D-MINT has a higher probability of RI compared to RIPA. For Southern Hemisphere TCs (Fig. 12e), D-MINT is also more skillful than RIPA for 20-kt RI. On average, D-MINT has a higher probability of 20-kt RI compared to RIPA, but a lower probability for all other thresholds. While this is similar to the NIO, the frequency of RI is lower in the Southern Hemisphere so it is unclear if these lower probabilities are contributing to lower skill. Overall, in these three basins, D-PRINT is less skillful than D-MINT, with the exception of 40-kt RI in Southern Hemisphere TCs.
d. Feature contributions to TC intensity
As noted earlier, SHAP values can be used to estimate which features in the satellite images increase or decrease the predicted TC intensity compared to the training dataset average. Figure 13 depicts IR features and SHAP values from Hurricane Iota (2020), which was chosen since it was a strong TC (MSW of 85 kt) that also underwent a 45-kt RI in 24 h. In Fig. 13, the satellite IR images are displayed in the top row with corresponding SHAP ordered vertically by lead time. If no image is displayed, no IR image of that age is used for predicting TC intensity for that lead time. For current TC intensity (Fig. 13b), there are relatively few contributing features. The presence of an eye in the 0-h IR image slightly increases the TC intensity, but not as noticeably as in stronger TCs (not shown). However, altogether the IR features increase the estimated TC intensity by 11.8 kt. For the 6-h forecast lead time (Fig. 13c), two different characteristics are apparent. First, the eye characteristic in the 0-h IR image positively influences the predicted intensity, consistent with how TC eye formation signifies intensification (Knaff and DeMaria 2017). Yet, the 6-h IR image has the largest IR contribution to the predicted TC intensity. This contribution is possibly due to the decrease in BTs in the eyewall region between the 6- and 0-h IR images, as eyewall formation is also indicative of an intensifying TC (Willoughby 1979; Shapiro and Willoughby 1982). While D-MINT does not explicitly use an IR difference feature like AI-RI (Griffin et al. 2022), D-MINT can infer these differences internally. SHAP values for 12- and 18-h lead times are also highest in the eye and eyewall region, consistent with research indicating that eyewall convection increases TC intensity (Jiang 2012). The impact of the eyewall on predicted TC intensity is decreased in the SHAP values for a 24-h lead time, when D-MINT also predicts Hurricane Iota’s intensity will decrease.
IR imagery from Hurricane Iota (2020) and corresponding SHAP values. If there is no SHAP value displayed, that IR image is not used for predicting TC intensity at that lead time.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
Figure 14 depicts the SHAP values for the MW features for Hurricane Iota. In Fig. 14a, an eyewall feature is identified by the lower BTs in the 89-GHz image (right). This eyewall feature positively influences the estimated current intensity, consistent with Wimmers et al. (2019), as well as the predicted future intensity. The occurrence of these lower BTs has been associated with TC intensification (Alvey et al. 2015), as lower BTs in 89-GHz are indicative of convection. The 37-GHz image (left) also positively influences the predicted intensity due to enhanced BTs around the center of the TC, a feature consistent with RI occurring (Fischer et al. 2018; Tao and Jiang 2015; Harnos and Nesbitt 2016). These warmer BTs indicate increased liquid water path from either shallow convection or deep convection without appreciable ice content.
MW imagery from Hurricane Iota (2020) and corresponding SHAP values.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
Figures 13 and 14 depict the SHAP values for a strong and intensifying TC. An example of SHAP values from a weaker TC, Tropical Storm Jerry with a MSW of 55 kt with no intensification over the next 24 h, are shown in the online supplemental figures. In this case, IR and MW features corresponding to positive and negative SHAP values are not as easily associated with satellite features compared to the well-organized TC Iota. However, TC Jerry (2019) is a good example of SHAP values when only a partial MW image is available. Since D-MINT cannot accept NaN (not a number) values for missing imagery, missing imagery is given a normalized value of 0. As a result, the missing imagery is still subject to SHAP calculations, although the SHAP values will usually reflect the low importance of the missing image section.
To shed more light on the impact of features to the predicted TC intensity, SHAP values for all 2019–21 Global TCs and for all forecast lead times are shown in Fig. 15. Here the SHAP values for each forecast lead time are arrayed in a single row, with the current intensity analysis at the top and 24-h forecast at the bottom (see inset in lower right corner). Some IR features, such as 12h_old_IR, are only used when predicting current intensity, and thus only one row of SHAP values is displayed. Each forecast is represented by a dot, with the color of each dot indicating its corresponding normalized value as input data. Based on Fig. 15, satellite features have the largest range of impact on the predicted TC intensity. Some trends can be identified in the IR satellite features, such as higher feature values (lower BTs) increasing the predicted intensity. Lower values in 37-GHz MW images, which are higher BTs (see normalization in Fig. 1b), are associated with increased SHAP values. SHAP values for 89-GHz MW imagery do not show a distinguishable trend, as the highest values can either increase or decrease the predicted TC intensity. When high values of 89-GHz MW imagery decrease the predicted TC intensity, this is associated with weaker TCs and disorganized convection.
Feature SHAP values for 2019–21 global TCs. SHAP values for each forecast lead time are in a single row, with the current intensity analysis at the top and 24-h forecast at the bottom (see inset in bottom-right corner). Each forecast is represented by a dot, with the color of each individual dot indicating the value of a given feature is low (blue) or high (red) compared to all values of that given feature in 2019–21 global TCs.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-23-0085.1
For scalar features, clear trends can be identified between feature values and impact on predicted TC intensity. Some of these trends are intuitive and consistent with Griffin et al. (2022). The scalar feature with the largest range is DELV, with the red dots associated with larger SHAP values indicating that a higher rate of intensification in the previous 12 h is correlated with an increase in the predicted intensity. Warmer underlying ocean values (RSST, CD20, CD26) increase the predicted intensity, while higher vertical wind shear features (SHDC, SHRD, SHRS) have a decreasing influence. Other trends are also consistent with Griffin et al. (2022). High EPSS and low ENSS increase the predicted TC intensity. Some trends in scalar features are less intuitive or opposite from those observed in Griffin et al. (2022). While higher values of MPI and COHC increase the predicted TC intensity for the 24-h forecast, similar to Griffin et al. (2022), lower values of MPI and COHC are associated with an increase in TC intensity for forecast leads times of 12 h or less. One explanation for this trend is COHC and MPI decrease with increasing latitude. TCs reach a maximum intensity around 20° latitude (Kossin et al. 2014), where the COHC is lower than observed in the tropics. Therefore, D-MINT is possibly identifying lower COHC and MPI as indicative of a stronger than average TC. This explanation also explains why sin_lat increases the predicted TC intensity. Another SHAP value trend differing from Griffin et al. (2022) is the tangential wind predictors (V850, V500). Since stronger TCs have a higher tangential wind (Doyle et al. 2017; Stern and Nolan 2011), D-MINT is indicating higher values are associated with an increased TC intensity.
6. Summary and conclusions
This study develops two convolutional neural network models, named D-MINT and D-PRINT, to predict current and short-term intensity of global TCs using two-dimensional features in satellite infrared (IR) and passive microwave (MW) imagery, as well as scalar environmental features. The difference between the two models is while D-PRINT only examines IR imagery, D-MINT also includes information-rich MW imagery (albeit with much less frequent availability). Optimal feature selection is determined by first developing a version of D-MINT that uses all scalar features listed in Table 1 in addition to the satellite features, and then reordering the data for one feature at a time and recalculating the predicted intensity for lead times of 6, 12, 18, and 24 h to determine the feature’s impact on TC intensity prediction skill. A feature is considered optimal if the predicted intensity RMSE increases with the feature’s randomization. These optimized scalar predictors are then used for all subsequent D-MINT and D-PRINT predictions of TC intensity at 0, 6, 12, 18, and 24 h. D-MINT and D-PRINT are trained on global TCs from 1994 to 2015.
D-MINT and D-PRINT are tested on an independent dataset consisting of global TCs from 2019 to 2021. Estimates of current TC intensity are compared to three established satellite-based estimates: ADT, AiDT, and SATCON. Performance results indicate that for the global sample as a whole, as well as for the North Atlantic, and eastern and western North Pacific TC basins, D-MINT and D-PRINT are more skillful than the ADT, and D-MINT is more skillful than the AiDT and SATCON. D-MINT and D-PRINT are also more skillful than ADT in the north Indian Ocean and Southern Hemisphere, but less skillful than SATCON and AiDT in these basins, possibly due to the relative lack of weaker TCs in these datasets.
D-MINT/D-PRINT short-term deterministic TC future intensity predictions are compared to five sources of operational forecasts: the official NHC or JTWC forecast, SHIPS/LGEM/DSHIPS (whichever has highest skill), HWRF, HMON, and ICON (latter two only for North Atlantic and eastern North Pacific TCs). Overall, D-MINT is more skillful than D-PRINT. In the North Atlantic and eastern North Pacific, D-MINT and D-PRINT are less skillful than the official NHC forecast and ICON, but D-MINT is generally more skillful than the other models. In the western North Pacific and north Indian Ocean, D-MINT has a lower predicted intensity RMSE than all other forecast models and the official JTWC forecasts, and D-PRINT has a lower RMSE than all but the official JTWC forecasts for north Indian Ocean TCs. For Southern Hemisphere TCs, D-MINT has a lower RMSE than the official forecast for 12-, 18-, and 24-h forecast lead times, and a lower RMSE than LGEM (except 6-h forecast) and HWRF. D-PRINT is less skillful than the official JTWC forecast, and is only more skillful than LGEM and HWRF for 18- and 24-h forecasts.
Probabilistic predictions of RI for five thresholds, 20 kt in 12 h and 25-, 30-, 35-, and 40-kt in 24 h, are compared to four models: SHIPS-Consensus, DTOPS, and AI-RI in North Atlantic and eastern North Pacific TCs and RIPA for other basins. For North Atlantic TCs, D-MINT is the most skillful at predicting 20-, 25-, and 30-kt RI. For eastern North Pacific TCs, DTOPS is the most skillful, but D-MINT and D-PRINT are more skillful than DTOPS for 24-h RI in TCs north of 15°N. In the western North Pacific, D-MINT is more skillful than RIPA for predicting all 24-h RI thresholds, but is less skillful for north Indian Ocean TCs. For 20-kt RI, D-MINT is more (less) skillful than RIPA in the north Indian Ocean (western North Pacific). For Southern Hemisphere TCs, D-MINT is more skillful than RIPA for 20- and 40-kt RI.
In addition to showing skill in predicting TC intensity, D-MINT and D-PRINT highlight the importance of interrogating two-dimensional multispectral satellite imagery in combination with environmental features. Many current intensity estimation methods rely solely on satellite imagery, but the addition of environmental features can increase skill. Also, trends in satellite images are also not considered by many of the current methods, but can have the highest predictor impact in D-MINT and D-PRINT. Many operational forecast models, like SHIPS and LGEM, reduce IR satellite imagery to scalar predictors, which can struggle to quantify features like convection. Therefore, satellite predictors may have less weight in those operational models than environmental predictors. The convolutional neural network used here is more appropriate for this task, and thus satellite predictors in this study have the highest SHAP values. This importance occurs even as satellite feature SHAP values decrease with increasing forecast lead time, and the importance of environmental features increases. In addition, these operational forecast models generally do not directly include satellite MW data, which though sporadic, can also improve intensity prediction. Jones et al. (2006) did note that including scalar features from MW imagery improved SHIPS intensity forecasts.
There are many different ideas for future work on D-MINT and D-PRINT. To determine whether the impact of satellite images on D-MINT/D-PRINT intensity forecasts is greater due to their two-dimensional nature, D-MINT and D-PRINT will be trained with two-dimensional environmental features, such as vertical wind shear and layered precipitable water, to identify whether the location/orientation of these features with respect to the TC core can improve intensity prediction. This will also and empirically demonstrate the relative importance of two- and three-dimensional interactions between the TC core and its environment. To further assess the impact of the orientation of TCs, D-MINT/D-PRINT will be retrained with flipped SH TCs to mimic TCs in the Northern Hemisphere. To quantify the impact of selecting the same scalar predictors and IR hours for both D-MINT and D-PRINT, optimal scalar predictors will be identified with for each combination of forecast lead time and IR hours, and the optimal model configuration will be selected for D-MINT and D-PRINT individually for each forecast lead time. Finally, additional future work will include exploring recurrent neural networks to predict TC intensity given their ability to process temporal information, since satellite trends often yield important information.
Acknowledgments.
The authors of this paper would like to thank Mark DeMaria and the SHIPS team at the Regional and Mesoscale Meteorology Branch for providing the archived SHIPS predictor files used to obtain the scalar features. Also, thank you to Ryan Lagerquist, John Cintineo, and Charles White for their assistance in optimizing the process to develop D-MINT and D-PRINT. This research was funded by the Office of Naval Research Award N00014-20-1-2149: A Deep Learning Approach to Examining and Predicting TC Rapid Intensification. Thank you to Kim Wood and an anonymous reviewer for their suggestions that greatly improved the manuscript.
Data availability statement.
Tropical cyclone best tracks can be found at https://ftp.nhc.noaa.gov/atcf/archive/ and https://www.metoc.navy.mil/jtwc/jtwc.html?best-tracks. Archived satellite data can be found via the Space Science and Engineering online archive. More information can be found at https://www.ssec.wisc.edu/datacenter/goes-archive/. The microwave imagery from NRL TC (MINT) dataset is available upon request from the corresponding author.
REFERENCES
Alvey, G. R., III, J. Zawislak, and E. Zipser, 2015: Precipitation properties observed during tropical cyclone intensity change. Mon. Wea. Rev., 143, 4476–4492, https://doi.org/10.1175/MWR-D-15-0065.1.
Cangialosi, J. P., E. Blake, M. DeMaria, A. Penny, A. Latto, E. Rappaport, and V. Tallapragada, 2020: Recent progress in tropical cyclone intensity forecasting at the National Hurricane Center. Wea. Forecasting, 35, 1913–1922, https://doi.org/10.1175/WAF-D-20-0059.1.
Chen, B.-F., B. Chen, H.-T. Lin, and R. L. Elsberry, 2019: Estimating tropical cyclone intensity by satellite imagery utilizing convolutional neural networks. Wea. Forecasting, 34, 447–465, https://doi.org/10.1175/WAF-D-18-0136.1.
Cossuth, J., S. Yang, K. Richardson, M. Surratt, J. Solbrig, and J. Hawkins, 2013: Creating a consistent climatology of tropical cyclone structure as observed by satellite microwave sensors. Special Symp. on the Next Level of Predictions in Tropical Meteorology: Techniques, Usage, Support, and Impacts, Austin, TX, Amer. Meteor. Soc., TJ25.5, https://ams.confex.com/ams/93Annual/webprogram/Paper220790.html.
DeMaria, M., 2009: A simplified dynamical system for tropical cyclone intensity prediction. Mon. Wea. Rev., 137, 68–82, https://doi.org/10.1175/2008MWR2513.1.
DeMaria, M., and J. Kaplan, 1994: A Statistical Hurricane Intensity Prediction Scheme (SHIPS) for the Atlantic basin. Wea. Forecasting, 9, 209–220, https://doi.org/10.1175/1520-0434(1994)009<0209:ASHIPS>2.0.CO;2.
DeMaria, M., M. Mainelli, L. K. Shay, J. A. Knaff, and J. Kaplan, 2005: Further improvements to the Statistical Hurricane Intensity Prediction Scheme (SHIPS). Wea. Forecasting, 20, 531–543, https://doi.org/10.1175/WAF862.1.
DeMaria, M., J. L. Franklin, M. J. Onderlinde, and J. Kaplan, 2021: Operational forecasting of tropical cyclone rapid intensification at the National Hurricane Center. Atmosphere, 12, 683, https://doi.org/10.3390/atmos12060683.
Doyle, J. D., and Coauthors, 2017: A view of tropical cyclones from above: The tropical cyclone intensity experiment. Bull. Amer. Meteor. Soc., 98, 2113–2134, https://doi.org/10.1175/BAMS-D-16-0055.1.
Draper, D. W., D. A. Newell, F. J. Wentz, S. Krimchansky, and G. M. Skofronick-Jackson, 2015: The Global Precipitation Measurement (GPM) Microwave Imager (GMI): Instrument overview and early on-orbit performance. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 8, 3452–3462, https://doi.org/10.1109/JSTARS.2015.2403303.
Emanuel, K., and F. Zhang, 2016: On the predictability and error sources of tropical cyclone intensity forecasts. J. Atmos. Sci., 73, 3739–3747, https://doi.org/10.1175/JAS-D-16-0100.1.
Fetanat, G., A. Homaifar, and K. R. Knapp, 2013: Objective tropical cyclone intensity estimation using analogs of spatial features in satellite data. Wea. Forecasting, 28, 1446–1459, https://doi.org/10.1175/WAF-D-13-00006.1.
Fischer, M. S., B. H. Tang, K. L. Corbosiero, and C. M. Rozoff, 2018: Normalized convective characteristics of tropical cyclone rapid intensification events in the North Atlantic and eastern North Pacific. Mon. Wea. Rev., 146, 1133–1155, https://doi.org/10.1175/MWR-D-17-0239.1.
Global Change Observing Mission, 2011: Appendix C Overview of the Global Change Observation Mission (GCOM). JAXA, 12 pp., https://suzaku.eorc.jaxa.jp/GCOM_W/w_amsr2/GCOM_RA_3rd_Guide_AppendixC.pdf.
Griffin, S. M., A. Wimmers, and C. S. Velden, 2022: Predicting rapid intensification in North Atlantic and eastern North Pacific tropical cyclones using a convolutional neural network. Wea. Forecasting, 37, 1333–1355, https://doi.org/10.1175/WAF-D-21-0194.1.
Harnos, D. S., and S. W. Nesbitt, 2016: Passive microwave quantification of tropical cyclone inner-core cloud populations relative to subsequent intensity change. Mon. Wea. Rev., 144, 4461–4482, https://doi.org/10.1175/MWR-D-15-0090.1.
Hinton, G. E., N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov, 2012: Improving neural networks by preventing co-adaptation of feature detectors. arXiv, 1207.0580v1, https://arxiv.org/pdf/1207.0580.pdf.
Huang, X., X. Peng, J. Fei, X. Cheng, J. Ding, and D. Yu, 2021: Evaluation and error analysis of official tropical cyclone intensity forecasts during 2005–2018 for the western North Pacific. J. Meteor. Soc. Japan, 99, 139–163, https://doi.org/10.2151/jmsj.2021-008.
Jergensen, G. E., A. McGovern, R. Lagerquist, and T. Smith, 2020: Classifying convective storms using machine learning. Wea. Forecasting, 35, 537–559, https://doi.org/10.1175/WAF-D-19-0170.1.
Jiang, H., 2012: The relationship between tropical cyclone intensity change and the strength of inner-core convection. Mon. Wea. Rev., 140, 1164–1176, https://doi.org/10.1175/MWR-D-11-00134.1.
Jiang, H., C. Tao, and Y. Pei, 2019: Estimation of tropical cyclone intensity in the North Atlantic and northeastern Pacific basins using TRMM satellite passive microwave observations. J. Appl. Meteor. Climatol., 58, 185–197, https://doi.org/10.1175/JAMC-D-18-0094.1.
Jones, T. A., D. J. Cecil, and M. DeMaria, 2006: Passive microwave-enhanced statistical hurricane intensity prediction scheme. Wea. Forecasting, 21, 613–635, https://doi.org/10.1175/WAF941.1.
Kaplan, J., and Coauthors, 2015: Evaluating environmental impacts on tropical cyclone rapid intensification predictability utilizing statistical models. Wea. Forecasting, 30, 1374–1396, https://doi.org/10.1175/WAF-D-15-0032.1.
Kawanishi, T., and Coauthors, 2003: The Advanced Microwave Scanning Radiometer for the Earth observing system (AMSR-E), NASDA’s contribution to the EOS for global energy and water cycle studies. IEEE Trans. Geosci. Remote Sens., 41, 184–194, https://doi.org/10.1109/TGRS.2002.808331.
Knaff, J. A., and R. T. DeMaria, 2017: Forecasting tropical cyclone eye formation and dissipation in infrared imagery. Wea. Forecasting, 32, 2103–2116, https://doi.org/10.1175/WAF-D-17-0037.1.
Knaff, J. A., C. R. Sampson, and K. D. Musgrave, 2018: An operational rapid intensification prediction aid for the western North Pacific. Wea. Forecasting, 33, 799–811, https://doi.org/10.1175/WAF-D-18-0012.1.
Knaff, J. A., C. R. Sampson, and B. R. Strahl, 2020: A tropical cyclone rapid intensification prediction aid for the joint typhoon warning center’s areas of responsibility. Wea. Forecasting, 35, 1173–1185, https://doi.org/10.1175/WAF-D-19-0228.1.
Kossin, J. P., K. A. Emanuel, and G. A. Vecchi, 2014: The poleward migration of the location of tropical cyclone maximum intensity. Nature, 509, 349–352, https://doi.org/10.1038/nature13278.
Kummerow, C., W. Barnes, T. Kozu, J. Shiue, and J. Simpson, 1998: The Tropical Rainfall Measuring Mission (TRMM) sensor package. J. Atmos. Oceanic Technol., 15, 809–817, https://doi.org/10.1175/1520-0426(1998)015<0809:TTRMMT>2.0.CO;2.
Kunkee, D. B., G. A. Poe, D. J. Boucher, S. D. Swadley, Y. Hong, J. E. Wessel, and E. A. Uliana, 2008: Design and evaluation of the first Special Sensor Microwave Imager/Sounder (SSMIS). IEEE Trans. Geosci. Remote Sens., 46, 863–883, https://doi.org/10.1109/TGRS.2008.917980.
Lagerquist, R., A. McGovern, and D. J. Gagne II, 2019: Deep learning for spatially explicit prediction of synoptic-scale fronts. Wea. Forecasting, 34, 1137–1160, https://doi.org/10.1175/WAF-D-18-0183.1.
Lee, J., J. Im, D.-H. Cha, H. Park, and S. Sim, 2020: Tropical cyclone intensity estimation using multi-dimensional convolutional neural networks from geostationary satellite data. Remote Sens., 12, 108, https://doi.org/10.3390/rs12010108.
Lundberg, S. M., and S.-I. Lee, 2017: A unified approach to interpreting model predictions. arXiv, 1705.07874v2, http://arxiv.org/abs/1705.07874.
Lundberg, S. M., G. G. Erion, and S.-I. Lee, 2018: Consistent individualized feature attribution for tree ensembles. arXiv, 1802.03888v3, http://arxiv.org/abs/1802.03888.
Lundberg, S. M., and Coauthors, 2020: From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell., 2, 56–67, https://doi.org/10.1038/s42256-019-0138-9.
Maas, A. L., A. Y. Hannun, and A. Y. Ng, 2013: Rectifier nonlinearities improve neural network acoustic models. Proc. 30th Int. Conf. on Machine Learning, Atlanta, GA, International Machine Learning Society, 6 pp., https://ai.stanford.edu/∼amaas/papers/relu_hybrid_icml2013_final.pdf.
Mangalathu, S., S.-H. Hwang, and J.-S. Jeon, 2020: Failure mode and effects analysis of RC members based on machine-learning-based SHapley Additive exPlanations (SHAP) approach. Eng. Struct., 219, 110927, https://doi.org/10.1016/j.engstruct.2020.110927.
Mehra, A., V. Tallapragada, Z. Zhang, B. Liu, L. Zhu, W. Wang, and H.-S. Kim, 2018: Advancing the state of the art in operational tropical cyclone forecasting at NCEP. Trop. Cyclone Res. Rev., 7, 51–56, https://doi.org/10.6057/2018TCRR01.06.
Olander, T. L., and C. S. Velden, 2019: The Advanced Dvorak Technique (ADT) for estimating tropical cyclone intensity: Update and new capabilities. Wea. Forecasting, 34, 905–922, https://doi.org/10.1175/WAF-D-19-0007.1.
Olander, T. L., A. Wimmers, C. Velden, and J. P. Kossin, 2021: Investigation of machine learning using satellite-based advanced Dvorak technique analysis parameters to estimate tropical cyclone intensity. Wea. Forecasting, 36, 2161–2186, https://doi.org/10.1175/WAF-D-20-0234.1.
Onderlinde, M., and M. DeMaria, 2018: Deterministic to Probabilistic Statistical rapid intensification index (DTOPS): A new method for forecasting RI probability. 33rd Conf. on Hurricanes and Tropical Meteorology, Ponte Vedra, FL, Amer. Meteor. Soc., 16C.3, https://ams.confex.com/ams/33HURRICANE/webprogram/Paper339346.html.
Pan, B., X. Xu, and Z. Shi, 2019: Tropical cyclone intensity prediction based on recurrent neural networks. Electron. Lett., 55, 413–415, https://doi.org/10.1049/el.2018.8178.
Raytheon Systems Company, 2000: SSM/I user’s interpretation guide. 104 pp., Tech. Doc. UG32268-900, https://gsics.atmos.umd.edu/pub/Development/MicrowaveImagers/SSMI_Users_Interpretation_Guide_NOV00.pdf.
Ritchie, E. A., G. Valliere-Kelley, M. F. Piñeros, and J. S. Tyo, 2012: Tropical cyclone intensity estimation in the North Atlantic basin using an improved deviation angle variance technique. Wea. Forecasting, 27, 1264–1277, https://doi.org/10.1175/WAF-D-11-00156.1.
Ritchie, E. A., K. M. Wood, O. G. Rodríguez-Herrera, M. F. Piñeros, and J. S. Tyo, 2014: Satellite-derived tropical cyclone intensity in the North Pacific Ocean using the deviation-angle variance technique. Wea. Forecasting, 29, 505–516, https://doi.org/10.1175/WAF-D-13-00133.1.
Rozoff, C. M., C. S. Velden, J. Kaplan, J. P. Kossin, and A. J. Wimmers, 2015: Improvements in the probabilistic prediction of tropical cyclone rapid intensification with passive microwave observations. Wea. Forecasting, 30, 1016–1038, https://doi.org/10.1175/WAF-D-14-00109.1.
Shapiro, L. J., and H. E. Willoughby, 1982: The response of balanced hurricanes to local sources of heat and momentum. J. Atmos. Sci., 39, 378–394, https://doi.org/10.1175/1520-0469(1982)039<0378:TROBHT>2.0.CO;2.
Shapley, L. S., 1953: A value for n-Person Games. Contribution to Theory Games (AM-28), H. W. Kuhn and A. W. Tucker, Eds., Princeton University Press, 307–318, https://doi.org/10.1515/9781400881970-018.
Stern, D. P., and D. S. Nolan, 2011: On the vertical decay of the maximum tangential winds in tropical cyclones. J. Atmos. Sci., 68, 2073–2094, https://doi.org/10.1175/2011JAS3682.1.
Surgi, N., 2008: Advanced Hurricane Prediction at NCEP’S Environmental Model Center: The Operational Implementation of the HWRF. NCEP, 32 pp.
Tao, C., and H. Jiang, 2015: Distributions of shallow to very deep precipitation–convection in rapidly intensifying tropical cyclones. J. Climate, 28, 8791–8824, https://doi.org/10.1175/JCLI-D-14-00448.1.
Velden, C. S., and D. Herndon, 2020: A consensus approach for estimating tropical cyclone intensity from meteorological satellites: SATCON. Wea. Forecasting, 35, 1645–1662, https://doi.org/10.1175/WAF-D-20-0015.1.
Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences. 2nd ed. International Geophysics Series, Vol. 100, Academic Press, 648 pp.
Willoughby, H. E., 1979: Excitation of spiral bands in hurricanes by interaction between the symmetric mean vortex and a shearing environmental steering current. J. Atmos. Sci., 36, 1226–1235, https://doi.org/10.1175/1520-0469(1979)036<1226:EOSBIH>2.0.CO;2.
Wimmers, A., C. Velden, and J. H. Cossuth, 2019: Using deep learning to estimate tropical cyclone intensity from satellite passive microwave imagery. Mon. Wea. Rev., 147, 2261–2282, https://doi.org/10.1175/MWR-D-18-0391.1.
Xu, W., K. Balaguru, A. August, N. Lalo, N. Hodas, M. DeMaria, and D. Judi, 2021: Deep learning experiments for tropical cyclone intensity forecasts. Wea. Forecasting, 36, 1453–1470, https://doi.org/10.1175/WAF-D-20-0104.1.
Yang, S., J. Hawkins, and K. Richardson, 2014: The improved NRL tropical cyclone monitoring system with a unified microwave brightness temperature calibration scheme. Remote Sens., 6, 4563–4581, https://doi.org/10.3390/rs6054563.
Zhao, Y., C. Zhao, R. Sun, and Z. Wang, 2016: A multiple linear regression model for tropical cyclone intensity estimation from satellite infrared images. Atmosphere, 7, 40, https://doi.org/10.3390/atmos7030040.
Zhuo, J.-Y., and Z.-M. Tan, 2021: Physics-augmented deep learning to improve tropical cyclone intensity and size estimation from satellite imagery. Mon. Wea. Rev., 149, 2097–2113, https://doi.org/10.1175/MWR-D-20-0333.1.
