Estimating Current Tropical Cyclone Intensity from 183-GHz Microwave Satellite Observations Using a Convolutional Neural Network

Recently, various artificial intelligence (AI)-based techniques have been developed to estimate tropical cyclone (TC) intensity by analyzing patterns in satellite imagery. Most of these models rely solely on currently available satellite data. However, traditional methods used in operational TC forecasting, such as the advanced Dvorak technique (ADT), as well as forecasters’ practices in TC intensity estimation, consider not only the current information but also historical data (e.g., previous TC intensities and their trends). Additionally, forecasters compare TC intensity from ADT and analysis track data, and if a bias is identified, they take this into account when estimating the current TC intensity. This study investigates how much the performance of current AI-based TC intensity estimation can be improved by incorporating historical information and bias estimates. Three experiments were conducted: (1) control (CTRL), which uses only current satellite images; (2) Experiment 1 (Exp 1), which includes current data along with satellite imagery and intensity from 6 hours earlier; and (3) Experiment 2 (Exp 2), which further incorporates bias estimates from the past 18 hours into a self-diagnosis algorithm. The results showed that Exp 1 reduced intensity estimation errors by 64.7% compared to CTRL, while Exp 2 further decreased errors by 32.3% compared to Exp 1. This suggests that utilizing historical TC information and a self-diagnosis approach in operational systems is crucial for improving the accuracy of AI-based TC intensity estimation.

Tropical cyclones (TCs) are among the most severe and destructive natural events, and they have a major detrimental impact on both the economy and society. This study used Geo-KOMPSAT-2A (GK2A) satellite data to estimate TC intensity in the western North Pacific based on a convolutional neural network (CNN) model. Given that the GK2A data cover only the period since 2019, we applied transfer learning to the model using information learned from previous Communication, Ocean, and Meteorological Satellite (COMS) data, which cover a considerably longer period (2011–2019). Transfer learning is a powerful technique that can improve the performance of a model even if the target task is based on a small amount of data. Experiments with various transfer learning methods using the GK2A and COMS data showed that the frozen–fine-tuning method had the best performance due to the high similarity between the two datasets. The test results for 2021 showed that employing transfer learning led to a 20% reduction in the root mean square error (RMSE) compared to models using only GK2A data. For the operational model, which additionally used TC images and intensities from six hours earlier, transfer learning reduced the RMSE by 5.5%. Because continuous long-term data are not always available for TC intensity estimation based on geostationary satellite images, these results imply that transfer learning may constitute a new advance in this area. Acknowledgement. This research was supported by Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Korea Coast Guard (RS-2023-00238652, Integrated Satellite-based Applications Development for Korea Coast Guard).  

The relationship between sea surface temperature (SST) and tropical cyclone (TC) intensity change exhibits a strong dependence on the current TC intensity. Using western North Pacific TC observations from 1982 to 2018, a threshold SST (TSST) is identified as the SST required to maintain TC intensity. TSST increases with TC intensity, with TCs intensifying and weakening when SST is higher and lower than TSST, respectively. Across the dataset, mean TC intensity change is proportional to the difference between SST and TSST. This study also formulates an equation to quantify TC intensity change using SST and current TC intensity, which replicates 99.46% of the mean observed TC intensity changes. This equation could serve as an alternative to the linear regression‐based relationship between SST and TC intensity change that is widely used in statistical‐dynamical intensity models, thereby improving intensity forecasts.

This study examines the potential limit in predicting tropical cyclone (TC) intensity under idealized conditions. Using the phase-space reconstruction method for TC intensity time series obtained from the CM1 idealized simulations, it is found that CM1 axisymmetric dynamics contain low-dimensional chaos at the maximum intensity equilibrium. Examination of several attractor invariants including the largest Lyapunov exponent, the Sugihara–May correlation, and the correlation dimension captures a consistent range of the chaotic attractor dimension between 4 and 5 for TC intensity at the maximum intensity equilibrium. In addition, the intensity error doubling time estimated from the largest Lyapunov exponent is roughly 1–3 h, which accords with the decay time obtained from the Sugihara–May correlation. Furthermore, the findings in this study reveal a relatively short TC intensity predictability limit for CM1, which is ∼3–9 h based on the maximum tangential wind but noticeably longer for the minimum central pressure (∼12–18 h) after reaching the mature stage. So long as the traditional metrics for TC intensity such as the maximum surface wind or the minimum central pressure is used for intensity forecast, our results support that TC intensity forecast errors will not be reduced indefinitely in any model, even in the absence of all model and observational errors. As such, the future improvement of TC intensity forecast should be based on different metrics beyond the absolute intensity errors that are currently used in real-time intensity verification. Significance Statement Using the phase-space reconstruction method for tropical cyclone (TC) intensity time series obtained from idealized axisymmetric simulations, we show that TC axisymmetric dynamics in CM1 possesses low-dimensional chaos at the maximum intensity equilibrium. This low-dimensional dynamics explains the long tradition of representing TC intensity by a few measures as in the current practice. The chaotic property of CM1 axisymmetric dynamics also suggests a relatively short predictability range for TC intensity at the maximum intensity equilibrium. The potential existence of low-dimensional chaos for TC intensity in CM1 idealized simulations as found in this study supports the use of different intensity verification metrics beyond the traditional absolute intensity errors currently used in operational model evaluation.

Climatological characteristics of simulated intense tropical cyclones (TCs) in the western North Pacific were explored with a 20-km-mesh atmospheric general circulation model (AGCM20) and a 5-km-mesh regional atmospheric nonhydrostatic model (ANHM5). From the AGCM20 climate runs, 34 intense TCs with a minimum central pressure (MCP) less than or equal to 900 hPa were sampled. Downscaling experiments were conducted with the ANHM5 for each intense TC simulated by the AGCM20. Only 23 developed into TCs with MCP ≤ 900 hPa. Most of the best-track TCs with an MCP ≤ 900 hPa underwent rapid intensification (RI) and attained maximum intensities south of 25°N. The AGCM20 simulated a similar number of intense TCs as the best-track datasets. However, the intense AGCM20 TCs tended to intensify longer and more gradually; only half of them underwent RI. The prolonged gradual intensification resulted in significant northward shifts of the location of maximum intensity compared with the location derived from two best-track datasets. The inner-core structure of AGCM20 TCs exhibited weak and shallow eyewall updrafts with maxima below an altitude of 6 km, while downscaling experiments revealed that most of the intense ANHM5 TCs underwent RI with deep and intense eyewall updrafts and attained their maximum intensity at lower latitudes. The altitudes of updraft maxima simulated by the AGCM20 descended rapidly during the phase of greatest intensification as midlevel warming markedly developed. The change in major processes responsible for precipitation in AGCM20 TCs before and after maximum intensification suggests close relationships between the large-scale cloud scheme and midlevel warming and prolonged gradual intensification.

Tropical cyclone (TC) intensity is controlled by both environmental factors and its internal structure. This paper investigates the influence of inner core features derived from satellite data on TC intensity changes and explores effective methods to improve TC intensity forecast accuracy. We construct models for forecasting TC intensity over the next 5 days using the light gradient boosting machine (LGBM) algorithm in the northwestern Pacific. The models integrate climatological and persistence predictors, large-scale environmental variables at the initial time, inner core features extracted from satellite data, and variables related to the forecasted TC location. The models are trained on TC data from 2005 to 2019, tested using best track data from 2020 to 2022 and real-time data in 2023. Results show that the inner core features derived from satellite data are helpful for intensity forecasts within 36 h, and variables related to the forecasted location are helpful for improving TC intensity forecast accuracy in most cases, especially for the forecasts beyond 48 h. Compared with the National Centers for Environmental Prediction (NCEP) and European Centre for Medium-Range Weather Forecasts (ECMWF) models, our models demonstrate 4%–48% and 19%–70% decreases in mean absolute errors (MAEs) in real-time intensity forecasts. The most important predictors identified include potential future intensity change, intensity changes during the previous 12 h, vertical wind shear, and sea–land ratio. The inner core feature extracted from satellite data is also among the top 10 important variables for short-time intensity forecasts. Significance Statement This study advances tropical cyclone (TC) intensity forecasting by integrating climatological and persistence predictors, large-scale environmental variables at the initial time, satellite-derived inner core features, and variables related to the forecasted TC location into the light gradient boosting machine (LGBM) models. Our models significantly improve forecast accuracy, reducing mean absolute errors by 4%–48% compared with the National Centers for Environmental Prediction (NCEP) and 19%–70% compared with the European Centre for Medium-Range Weather Forecasts (ECMWF). Results denote that the inner core features improve the intensity prediction in the short time, and variables related to the forecasted TC location improve the intensity forecasts in most cases, especially for the forecasts beyond 48 h. This study provides valuable insights for enhancing operational TC intensity forecasting.

In this study, the relationship between tropical cyclone (TC) intensity and rain rate over the ocean is investigated using a full-physics numerical model (WRF) and a physics-based TC rainfall model (TCR). TC intensity is found to be nearly linearly correlated with the average rain rate in the inner core [∼0.97 (mm h−1 m−2)/(m s−1)], while the correlation is weak at outer radii. This difference is induced because TC intensity is significantly correlated with both the vertical velocity and specific humidity in the inner core but is not significantly correlated with the vertical velocity in the outer radii. Further investigation shows that the intensity–rain-rate relationship at the outer radii is influenced by the TC evolution stage. The rain rate for the outer radii is positively correlated with TC intensity for nondecaying TCs, while this correlation is reduced for decaying TCs due to systematic downdrafts in the outer radii. In the context of climate change, the sensitivity of the TC rain rate to sea surface temperature (SST) is found to be +9% per 1 K increase of SST, roughly the product of the sensitivity of TC intensity to SST (+3%) and the Clausius-Clapeyron scaling (+7%). Coupled with synthetic storms, evolution of the TC rain rate over the twenty-first century under the SSP5-8.5 scenario is projected by the TCR (calibrated with the WRF simulations). The annual increase rates of averaged TC rain rate are 0.17% and 0.20% for the inner core and outer radii, respectively, larger than the annual increase rate of TC intensity (0.046%) but comparable to that of cube of intensity (0.18%).

This study develops a probabilistic model based on a convolutional neural network to predict rapid intensification (RI) in both North Atlantic and eastern North Pacific tropical cyclones (TCs). Coined “I-RI,” an advantage of using a convolutional neural network to predict RI is that it is designed to learn from spatial fields, like two-dimensional satellite imagery, as well as scalar features. The resulting model RI probability output is validated against two operational RI guidances—an empirical and a deterministic method—to assess skill at predicting RI over 12-, 24-, 36-, 48-, and 72-h lead times. Results indicate that in North Atlantic TCs, AI-RI is more skillful at predicting RI over 12- and 24-h lead times compared to both operational RI guidances. In eastern North Pacific TCs, AI-RI is more skillful than the empirical operational RI guidance at most RI thresholds, but less skillful than the deterministic RI guidance at all thresholds. For TCs north of 15°N, where the deterministic skill was lower, AI-RI was more skillful than the deterministic operational guidance for over half of the RI thresholds. It is also found that AI-RI struggles to reach the higher RI probabilities produced by both of the operational RI guidances in both basins. This work demonstrates that the two-dimensional structures within the satellite imagery of TCs and the evolution of these structures identified using the difference in satellite images, captured by a convolutional neural network, yield better 12–24-h indicators of RI than existing scalar assessments of satellite brightness temperature. Significance Statement The purpose of this study is to develop a method to predict tropical cyclone rapid intensification using artificial intelligence. The developed model uses a convolutional neural network, which can identify features in satellite imagery that are indicative of rapid intensification. The results suggest that, compared with current operational rapid intensification models, a convolutional neural network approach is generally more skillful at predicting rapid intensification.

Estimating tropical cyclone (TC) intensity is crucial for disaster reduction and risk management. This study aims to estimate TC intensity using machine learning (ML) models. We utilized eight ML models to predict TC intensity, incorporating factors such as TC location, central pressure, distance to land, landfall in the next six hours, storm speed, storm direction, date, and number from the International Best Track Archive for Climate Stewardship Version 4 (IBTrACS V4). The dataset was divided into four sub-datasets based on the El Niño–Southern Oscillation (ENSO) phases (Neutral, El Niño, and La Niña). Our results highlight that central pressure has the greatest effect on TC intensity estimation, with a maximum root mean square error (RMSE) of 1.289 knots (equivalent to 0.663 m/s). Cubist and Random Forest (RF) models consistently outperformed others, with Cubist showing superior performance in both training and testing datasets. The highest bias was observed in SVM models. Temporal analysis revealed the highest mean error in January and November, and the lowest in February. Errors during the Warm phase of ENSO were notably higher, especially in the South China Sea. Central pressure was identified as the most influential factor for TC intensity estimation, with further exploration of environmental features recommended for model robustness.

An objective method for estimating tropical cyclone (TC) intensity using historical hurricane satellite data (HURSAT) is developed and tested. This new method, referred to as feature analogs in satellite imagery (FASI), requires a TC's center location to extract azimuthal brightness temperature (BT) profiles from current imagery as well as BT profiles from imagery 6, 12, and 24 h prior. Instead of using regression techniques, the estimated TC intensity is determined from the 10 closest analogs to this TC based on the BT profiles using a k-nearest-neighbor algorithm. The FASI technique was trained and validated using intensity data from aircraft reconnaissance in the North Atlantic Ocean, where the data were restricted to include storms that are over water and south of 45°N. This subset comprised 2016 observations from 165 storms during 1988–2006. Several tests were implemented to statistically justify the FASI algorithm using n-fold cross validation. The resulting average mean absolute intensity error was 10.9 kt (50% of estimates are within 10 kt, 1 kt = 0.51 m s−1) or 8.4 mb (50% of estimates are within 8 mb); its accuracy is on par with other objective techniques. This approach has the potential to provide global TC intensity estimates that could augment intensity estimates made by other objective techniques.

Recent studies have shown that assimilating enhanced satellite-derived atmospheric motion vectors (AMVs) has improved mesoscale forecast of tropical cyclones (TC) track and intensity. The authors conduct data-denial experiments to understand where the TC analyses and forecasts benefit the most from the enhanced AMV information using an ensemble Kalman filter and the Weather Research and Forecasting Model. The Cooperative Institute for Meteorological Satellite Studies at the University of Wisconsin provides enhanced AMV datasets with higher density and temporal resolution using shorter-interval image triplets for the duration of Typhoon Sinlaku and Hurricane Ike (both 2008). These AMV datasets are then spatially and vertically subsetted to create six parallel cycled assimilation-forecast experiments for each TC: all AMVs; AMVs withheld between 100 and 350 hPa (upper layer), between 350 and 700 hPa (middle layer), and between 700 and 999 hPa (lower layer); and only AMVs within (interior) and outside (exterior) 1000-km radius of the TC center. All AMV subsets are found to be useful in some capacity. The interior and upper-layer AMVs are particularly crucial for improving initial TC position, intensity, and the three-dimensional wind structure along with their forecasts. Compared with denying interior or exterior AMVs, withholding AMVs in different tropospheric layers had less impact on TC intensity and size forecasts. The ensemble forecast is less certain (larger spread) in providing accurate TC track, intensity, and size when upper-layer AMVs or interior AMVs are withheld. This information could be useful to potential targeting scenarios, such as activating and focusing satellite rapid-scan operations, and decisions regarding observing system assessments and deployments.

Using primitive equations with adiabatic, non‐friction non‐environmental steering, the dynamical equilibrium features of tropical cyclones (TC) are investigated. The parameter, dI/dt, was obtained and its dynamical balance formulation describing TC intensity change. The results of the we discussed the characteristics of TC motion and intensity change in barotropic and baroclinic model showed that the stationary motion of TC possess “equivalent‐geostrophic equilibrium” and isolines of potential vorticity and E(E = φ + (u2 + v2)/2) coincide with streamlines. The TC intensity will be weakened by “ventilation flow” in “beta (β) gyres” and “shear gyres” from the positive (negative) wind vertical shear at high‐ (low‐) level. The positive (negative) vorticity vertical convection of high‐ (low‐) level strengthens TC intensity. Generally, the TC center has the tendency to drift poleward and there is a gradient (ascendent) direction of vorticity vertical convection at high‐ (low‐) level. The TC motion is connected with the asymmetry structure of vertical shear and vertical motion, the “γ effect” (γ = ∂β/∂y) that is one of the dynamical mechanisms of “ventilation flow” steering TC center motion.

A tropical cyclone intensity prediction model using conditional generative adversarial network

Several key issues in the simple time-dependent theories of tropical cyclone (TC) intensification developed in recent years remain, including the lack of a closure for the pressure dependence of saturation enthalpy at sea surface temperature (SST) under the eyewall and the definition of environmental conditions, such as the boundary layer enthalpy in TC environment and the TC outflow-layer temperature. In this study, some refinements to the most recent time-dependent theory of TC intensification have been accomplished to resolve those issues. The first is the construction of a functional relationship between the surface pressure under the eyewall and the TC intensity, which is derived using the cyclostrophic wind balance and calibrated using full-physics axisymmetric model simulations. The second is the definition of TC environment that explicitly includes the air–sea temperature difference. The third is the TC outflow-layer temperature parameterized as a linear function of SST based on global reanalysis data. With these refinements, the updated time-dependent theory becomes self-contained and can give both the intensity-dependent TC intensification rate (IR) and the maximum potential intensity (MPI) under given environmental thermodynamic conditions. It is shown that the pressure dependence of saturation enthalpy at SST can lead to an increase in the TC MPI and IR by about half of that induced by dissipative heating due to surface friction. Results also show that both MPI and IR increase with increasing SST, surface enthalpy exchange coefficient, environmental air–sea temperature difference, and decreasing environmental boundary layer relative humidity, but the maximum IR is insensitive to surface drag coefficient. Significance Statement A new advancement in the recent decade is the development of simple time-dependent theories of tropical cyclone (TC) intensification, which can provide quantitative understanding of TC intensity change. However, several key issues in these simple time-dependent theories remain, including the lack of a closure for the pressure dependence of saturation enthalpy at sea surface temperature under the eyewall and the definition of environmental conditions. These are resolved in this study with several refinements, which make the most recent time-dependent theory of TC intensification self-contained and practical.

This study develops an objective deep-learning-based model for tropical cyclone (TC) intensity estimation. The model’s basic structure is a convolutional neural network (CNN), which is a widely used technology in computer vision tasks. In order to optimize the model’s structure and to improve the feature extraction ability, both residual learning and attention mechanisms are embedded into the model. Five cloud products, including cloud optical thickness, cloud top temperature, cloud top height, cloud effective radius, and cloud type, which are level-2 products from the geostationary satellite Himawari-8, are used as the model training inputs. We sampled the cloud products under the 13 rotational angles of each TC to augment the training dataset. For the independent test data, the model shows improvement, with a relatively low RMSE of 4.06 m/s and a mean absolute error (MAE) of 3.23 m/s, which are comparable to the results seen in previous studies. Various cloud organization patterns, storm whirling patterns, and TC structures from the feature maps are presented to interpret the model training process. An analysis of the overestimated bias and underestimated bias shows that the model’s performance is highly affected by the initial cloud products. Moreover, several controlled experiments using other deep learning architectures demonstrate that our designed model is conducive to estimating TC intensity, thus providing insight into the forecasting of other TC metrics.