A spatially-coherent attribution framework for interpreting black-box tropical cyclone intensity forecasts

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.

The present study uses the nonlinear singular vector (NFSV) approach to identify the optimally-growing tendency perturbations of the Weather Research and Forecasting (WRF) model for tropical cyclone (TC) intensity forecasts. For nine selected TC cases, the NFSV-tendency perturbations of the WRF model, including components of potential temperature and/or moisture, are calculated when TC intensities are forecasted with a 24-hour lead time, and their respective potential temperature components are demonstrated to have more impact on the TC intensity forecasts. The perturbations coherently show barotropic structure around the central location of the TCs at the 24-hour lead time, and their dominant energies concentrate in the middle layers of the atmosphere. Moreover, such structures do not depend on TC intensities and subsequent development of the TC. The NFSV-tendency perturbations may indicate that the model uncertainty that is represented by tendency perturbations but associated with the inner-core of TCs, makes larger contributions to the TC intensity forecast uncertainty. Further analysis shows that the TC intensity forecast skill could be greatly improved as preferentially superimposing an appropriate tendency perturbation associated with the sensitivity of NFSVs to correct the model, even if using a WRF with coarse resolution.

<p>The present study uses the nonlinear singular vector (NFSV) approach to identify the optimally-growing tendency perturbations of the Weather Research and Forecasting (WRF) model for tropical cyclone (TC) intensity forecasts. For nine selected TC cases, the NFSV-tendency perturbations of the WRF model, including components of potential temperature and/or moisture, are calculated when TC intensities are forecasted with a 24-hour lead time, and their respective potential temperature components are demonstrated to have more impact on the TC intensity forecasts. The perturbations coherently show barotropic structure around the central location of the TCs at the 24-hour lead time, and their dominant energies concentrate in the middle layers of the atmosphere. Moreover, such structures do not depend on TC intensities and subsequent development of the TC. The NFSV-tendency perturbations may indicate that the model uncertainty that is represented by tendency perturbations but associated with the inner-core of TCs, makes larger contributions to the TC intensity forecast uncertainty. Further analysis shows that the TC intensity forecast skill could be greatly improved as preferentially superimposing an appropriate tendency perturbation associated with the sensitivity of NFSVs to correct the model, even if using a WRF with coarse resolution.</p><div> <div> </div> </div>

This study explores the potential impact of global navigation satellite system radio occultation (RO) data assimilation on the tropical cyclone (TC) intensity forecast over the western North Pacific. The forecast experiments are performed through a regional model for six TCs occurring in 2020. RO data are mainly obtained from the Constellation Observing System for Meteorology, Ionosphere, and Climate Mission II. The forecasts with and without assimilation of RO data are compared, and their difference is regarded as the impact of RO data on TC forecasts. Overall, the forecasts tend to underestimate the TC intensity relative to the best track data. Compared to the forecasts assimilating without RO data, forecasts assimilating with RO data improve the initial conditions and reduce the underestimation of TC intensity forecast by 13 kt and 16 hPa in subsequent forecasts. This intensity improvement is more significant for TCs developing in drier environments than those in moister environments. The main period of intensity increase is 48-24 h prior to TCs developing to the maximum intensity. The assimilation of RO data increases the moisture around the TC centers, especially at mid-levels (700-300 hPa). It also increases the low-level vorticity but reduces the mid-level vorticity around the TC centers. These characteristics favor TCs with stronger surface wind speed and lower sea surface pressure. In summary, this study highlights the positive contribution of RO data to TC intensity forecast and explores the potential mechanisms.

The impact of vertical grid-nesting on the tropical cyclone intensity and track forecast was investigated using the Weather Research and Forecast (WRF) version 3.8 and the initialization method of the Structure Adjustable Balanced Bogus Vortex (SABV). For a better resolution in the central part of the numerical domain, where the tropical cyclone of interest is located, a horizontal and vertical nesting technique was employed. Simulations of the tropical cyclone Sanba (16th in 2012) indicated that the vertical nesting had a weak impact on the cyclone intensity and little impact on the track forecast. Further experiments revealed that the performance of forecast was quite sensitive to the horizontal resolution, which is in agreement with previous studies. The improvement is due to the fact that horizontal resolution can improve forecasts not only on the tropical cyclone-scale but also for large-scale disturbances.

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.

Track, intensity, and, in some cases, size are usually used as separate evaluation parameters to assess numerical model performance on tropical cyclone (TC) forecasts. Such an individual-parameter evaluation approach often encounters contradictory skill assessments for different parameters, for instance, small track error with large intensity error and vice versa. In this study, an intensity-weighted hurricane track density function (IW-HTDF) is designed as a new approach to the integrated evaluation of TC track, intensity, and size forecasts. The sensitivity of the TC track density to TC wind radius was investigated by calculating the IW-HTDF with density functions defined by 1) asymmetric, 2) symmetric, and 3) constant wind radii. Using the best-track data as the benchmark, IW-HTDF provides a specific score value for a TC forecast validated for a specific date and time or duration. This new TC forecast evaluation approach provides a relatively concise, integrated skill score compared with multiple skill scores when track, intensity and size are evaluated separately. It should be noted that actual observations of TC size data are very limited and so are the estimations of TC size forecasts. Therefore, including TC size as a forecast evaluation parameter is exploratory at the present. The proposed integrated evaluation method for TC track, intensity, and size forecasts can be used for evaluating the track forecast alone or in combination with intensity and size parameters. As observations and forecasts of TC size become routine in the future, including TC size as a forecast skill assessment parameter will become more imperative.

Reducing tropical cyclone (TC) intensity forecast errors is a challenging task that has interested the operational forecasting and research community for decades. To address this, we developed a deep learning (DL)-based Multilayer Perceptron (MLP) TC intensity prediction model. The model was trained using the global Statistical Hurricane Intensity Prediction Scheme (SHIPS) predictors to forecast the change in TC maximum wind speed for the Atlantic Basin. In the first experiment, a 24-hour forecast period was considered. To overcome sample size limitations, we adopted a Leave One Year Out (LOYO) testing scheme, where a model is trained using data from all years except one and then evaluated on the year that is left out. When tested on 2010–2018 operational data using the LOYO scheme, the MLP outperformed other statistical-dynamical models by 9-20%. Additional independent tests in 2019 and 2020 were conducted to simulate real-time operational forecasts, where the MLP model again outperformed the statistical-dynamical models by 5-22% and achieved comparable results as HWFI. The MLP model also correctly predicted more rapid intensification events than all the four operational TC intensity models compared. In the second experiment, we developed a lightweight MLP for 6-hour intensity predictions. When coupled with a synthetic TC track model, the lightweight MLP generated realistic TC intensity distribution in the Atlantic Basin. Therefore, the MLP-based approach has the potential to improve operational TC intensity forecasts, and will also be a viable option for generating synthetic TCs for climate studies.

For typhoon warning centers, effective forecasting of tropical cyclone intensity is always required. The major difficulties and challenges in forecasting tropical cyclone intensity are the complex physical mechanism and the structure of tropical cyclones. The interaction between the tropical cyclone and its environment is also a complex process. In this paper, a model based on support vector machines is developed to yield the 12, 24, 36, 48, 72 h forecasts of tropical cyclone intensity. Furthermore, the forecasts resulting from the proposed model are compared with those from the Joint Typhoon Warning Center. Cross-validation tests are also applied to evaluate the accuracy and the robustness of the proposed model. The results confirm that the proposed model can provide accurate forecasts of tropical cyclone intensity, especially for a long lead-time. When the sample events are classified into five categories according to the Saffir-Simpson scale, the forecasts resulting from the proposed model have the best performance for events in categories 4 and 5. In addition, when a typhoon turns northward, although the water temperature drops rapidly, the proposed model still performs well. In conclusion, the proposed model is useful to improve the forecasts of tropical cyclones intensity.

The development of a simple statistical tropical cyclone (TC) intensity forecast model is described. The primary purpose of this model, called southern hemisphere five-day statistical typhoon intensity forecast scheme (SH ST5D), is to provide a skill/no-skill control forecast for verifying other TC intensity forecasts. However, it also provides useful and always-available forecasts of TC intensity in the southern hemisphere. The model is created by fitting an optimal combination of factors related to climatology and persistence (or CLIPER) using multiple linear regression. These CLIPER factors are determined from the best track tropical cyclone dataset produced by the United States of America’s Joint Typhoon Warning Center (JTWC) in the years 1980-2002. In 2004 the SH ST5D model became part of the operational suite of tropical cyclone intensity guidance run at JTWC. The forecasts from the model since that time have outperformed both climatology (i.e. a constant 65 knots or 33 ms -1 forecast) and the persistence of initial conditions in a statistically significant manner in independent testing during 2004-2007. This documentation is provided to promote the use of this model’s output and provide adequate background for the development of similar models.

Over the last 30 years tropical cyclone (TC) intensity forecasts, for various (yet somewhat puzzling) reasons, have not achieved near the level of improvement of the TC track forecast. Although TCs have been studied intensively throughout the twentieth century, the community has surprisingly little quantitative knowledge as to how these storms interact with their environments, particularly with respect to changes in core structure (Frank & Ritchie, 1999). Rogers et al. (2006) stated that the lack of skill in numerical forecast of TC intensity can be partly attributed to inadequate understanding of the physics of TCs and the way they interact with their environment. In fact, TC structure and intensity changes are affected by a large and complex array of physical processes that govern the inner core structure and the interaction between the storm and both the underlying ocean and its atmospheric environment (Wang & Wu, 2004). Among other issues cited, crude parameterizations, difficulties in treating multiscale interactions, and the uncertainties involved with initializing the model over areas with sparse data coverage have received substantial attention. In order to predict TC intensity, one of the important questions has been how to first accurately predict a TC’s maximum potential intensity (MPI). Despite the fact that various methods for predicting a storm’s MPI have been put forth, the failure of the NWP community to realistically forecast TC intensity largely lies in the fact that there are various unexplained processes keeping TCs from reaching their theoretical MPI. While the mechanisms involved are myriad, there are essentially two kinds that have been identified as having the largest impact on TC intensification: 1) internal dynamics and 2) external forcing from environmental flow. Below these two headings fall most TC intensity-related topics: vertical wind shear-induced asymmetries in the core region, the cooling of the sea surface due to oceanic upwelling under the eyewall region, the role of inner and outer rainbands, vortex Rossby waves (VRWs), embedded mesovortices, and eyewall cycles. Tropical cyclones often fail to reach their theoretical MPI because prominent MPI calculations use the basic assumption of TC axisymmetry (Camp & Montgomery, 2001), whereas TC structure is rarely symmetric, even in mature storms. While the tangential wind field and other TC features are axisymmetric, many significant features, such as VRWs, eyewall cycles, rainfall, convection, radial winds, and outer rainbands are often highly asymmetric attributes that impact TC intensity change; it is thus no surprise that nearly all TCs fail to reach their MPI.

This paper proposes a method for multi-model ensemble forecasting based on Bayesian model averaging (BMA), aiming to improve the accuracy of tropical cyclone (TC) intensity forecasts, especially forecasts of minimum surface pressure at the cyclone center (Pmin). The multi-model ensemble comprises three operational forecast models: the Global Forecast System (GFS) of NCEP, the Hurricane Weather Research and Forecasting (HWRF) models of NCEP, and the Integrated Forecasting System (IFS) of ECMWF. The mean of a predictive distribution is taken as the BMA forecast. In this investigation, bias correction of the minimum surface pressure was applied at each forecast lead time, and the distribution (or probability density function, PDF) of Pmin was used and transformed. Based on summer season forecasts for three years, we found that the intensity errors in TC forecast from the three models varied significantly. The HWRF had a much smaller intensity error for short lead-time forecasts. To demonstrate the proposed methodology, cross validation was implemented to ensure more efficient use of the sample data and more reliable testing. Comparative analysis shows that BMA for this three-model ensemble, after bias correction and distribution transformation, provided more accurate forecasts than did the best of the ensemble members (HWRF), with a 5%–7% decrease in root-mean-square error on average. BMA also outperformed the multi-model ensemble, and it produced “predictive variance” that represented the forecast uncertainty of the member models. In a word, the BMA method used in the multi-model ensemble forecasting was successful in TC intensity forecasts, and it has the potential to be applied to routine operational forecasting.

Tropical cyclone (TC) intensity forecasting is crucial for early disaster warning and emergency decision-making. Numerous researchers have explored deep-learning methods to address computational and post-processing issues in operational forecasting. Regrettably, they exhibit subpar long-term forecasting capabilities. We use two strategies to enhance long-term forecasting. (1) By enhancing the matching between TC intensity and spatial information, we can improve long-term forecasting performance. (2) Incorporating physical knowledge and physical constraints can help mitigate the accumulation of forecasting errors. To achieve the above strategies, we propose the VQLTI framework. VQLTI transfers the TC intensity information to a discrete latent space while retaining the spatial information differences, using large-scale spatial meteorological data as conditions. Furthermore, we leverage the forecast from the weather prediction model FengWu to provide additional physical knowledge for VQLTI. Additionally, we calculate the potential intensity (PI) to impose physical constraints on the latent variables. In the global long-term TC intensity forecasting, VQLTI achieves state-of-the-art results for the 24h to 120h, with the MSW (Maximum Sustained Wind) forecast error reduced by 35.65%-42.51% compared to ECMWF-IFS.

Tropical cyclone (TC) intensity forecasts are impacted by errors in atmosphere and ocean initial conditions and the model formulation, which motivates using an ensemble approach. This study evaluates the impact of uncertainty in atmospheric and oceanic initial conditions, as well as stochastic representations of the drag Cd and enthalphy Ck exchange coefficients on ensemble Advanced Hurricane WRF (AHW) TC intensity forecasts of multiple Atlantic TCs from 2008 to 2011. Each ensemble experiment is characterized by different combinations of either deterministic or ensemble atmospheric and/or oceanic initial conditions, as well as fixed or stochastic representations of Cd or Ck. Among those experiments with a single uncertainty source, atmospheric uncertainty produces the largest standard deviation in TC intensity. While ocean uncertainty leads to continuous growth in ensemble standard deviation, the ensemble standard deviation in the experiments with Cd and Ck uncertainty levels off by 48 h. Combining atmospheric and oceanic uncertainty leads to larger intensity standard deviation than atmosphere or ocean uncertainty alone and preferentially adds variability outside of the TC core. By contrast, combining Cd or Ck uncertainty with any other source leads to negligible increases in standard deviation, which is mainly due to the lack of spatial correlation in the exchange coefficient perturbations. All of the ensemble experiments are deficient in ensemble standard deviation; however, the experiments with combinations of uncertainty sources generally have an ensemble standard deviation closer to the ensemble-mean errors.

It is well known that global numerical model analyses and forecasts benefit from the routine assimilation of atmospheric motion vectors (AMVs) derived from meteorological satellites. Recent studies have also shown that the assimilation of enhanced (spatial and temporal) AMVs can benefit research-mode regional model forecasts of tropical cyclone track and intensity. In this study, the impact of direct assimilation of enhanced (higher resolution) AMV datasets in the NCEP operational Hurricane Weather Research and Forecasting Model (HWRF) system is investigated. Forecasts of Atlantic tropical cyclone track and intensity are examined for impact by inclusion of enhanced AMVs via direct data assimilation. Experiments are conducted for AMVs derived using two methodologies (“HERITAGE” and “GOES-R”), and also for varying levels of quality control in order to assess and inform the optimization of the AMV assimilation process. Results are presented for three selected Atlantic tropical cyclone events and compared to Control forecasts without the enhanced AMVs as well as the corresponding operational HWRF forecasts. The findings indicate that the direct assimilation of high-resolution AMVs has an overall modest positive impact on HWRF forecasts, but the impact magnitudes are dependent on the 1) availability of rapid scan imagery used to produce the AMVs, 2) AMV derivation approach, 3) level of quality control employed in the assimilation, and 4) vortex initialization procedure (including the degree to which unbalanced states are allowed to enter the model analyses).