Advancing forecasting capabilities: A contrastive learning model for forecasting tropical cyclone rapid intensification
Edited by Richard Rotunno, National Center for Atmospheric Research, Boulder, CO; received August 1, 2024; accepted December 6, 2024
Significance
Tropical cyclones (TCs), especially those that rapidly intensify (RI), pose a significant threat due to the uncertainty in forecasting them. To improve RI TC forecasting, which intensify by at least 13 m/s within 24 h, a RI TC forecasting model was developed. The model uses satellite, atmospheric, and oceanic data to forecast TC changes. Tested on 1,149 TC periods in the Northwest Pacific from 2020 to 2021, it achieved a 92.3% accuracy in forecasting RI TC and significantly reduced false alarms to 8.9% (3 times reduction compared to existing best deep learning methods). By addressing issues like sample imbalance and including structural features of TCs, our model greatly improves forecasting accuracy, offering a morereliable way to forecast these dangerous weather events.
Abstract
Tropical cyclones (TCs), particularly those that rapidly intensify (RI), pose a significant threat due to the uncertainty in forecasting them. RI TC periods, which intensify by at least 13 m/s within 24 h, remain challenging to forecast accurately. Existing models achieve a probability of detection (POD) of 82.6% and a false alarm rate (FARate) of 27.2%. To address this, we developed a contrastive-based RI TC forecasting (RITCF-contrastive) model, utilizing satellite infrared imagery alongside atmospheric and oceanic data. The RITCF-contrastive model was tested on 1,149 TC periods in the Northwest Pacific from 2020 to 2021, achieving a POD of 92.3% and a FARate of 8.9%. RITCF-contrastive improves on previous models by addressing sample imbalance and incorporating TC structural features, leading to a 11.7% improvement in POD and a 3 times reduction in FARate compared to existing deep learning methods. The RITCF-contrastive model not only enhances RI TC forecasting but also offers a unique approach to forecasting these dangerous weather events.
Sign up for PNAS alerts.
Get alerts for new articles, or get an alert when an article is cited.
Tropical cyclones (TCs) can cause serious disasters to human beings. Among them, rapid intensification (RI) TC periods are defined as the intensity change equal to or more than 13 m/s within 24 h (1–4). RI TC periods account for only 5% of the total TC periods (In this work, a “TC case” refers to one TC, while a “TC period” refers to a specific time period during a TC case.), but because they are difficult to forecast (5), they often pose a more serious threat. Forecasting RI TC periods requires the analysis of atmospheric and oceanic environmental factors, such as wind, relative humidity, air temperature, and sea surface temperature (SST), as well as the development processes of TCs, such as eyewall replacement, RI of deep convection, etc. However, due to the large range of scales involved in TC intensification that are difficult to accurately measure, and inaccuracies of turbulent and microphysical parameterizations in numerical models, the forecast of RI TCs remains significantly inaccurate (6–8), which leads to severe damages (9–12). Therefore, improving the RI TCs’ forecast accuracy is a long-standing challenge for meteorological agencies worldwide (13).
RI TCs forecast methods mainly include numerical models (14, 15), statistical models (16–18), and deep learning (DL) models (5, 19–21). Numerical models like the Hurricane Weather Research and Forecasting model (HWRF) have limited accuracy in RI TC forecasts. In the Hurricane Forecast Improvement ProgramTechnical Report by the National Oceanic and Atmospheric Administration (22), the HWRF results showed a 15% probability of detection (POD) for RI TCs. In contrast, statistical models like Statistical Hurricane Intensity Prediction Scheme (SHIPS) enhanced by including satellite data and environmental factors (23) serve as primary guidance for RI TC forecasts at the NHC (National Hurricane Center), USA. However, the SHIPS method had only 50% POD (21).
In recent years, DL has demonstrated significant potential in science research (24–27), including TC intensity and track forecasting and RI TCs forecasting (28, 29). Li et al. (30) and Chandra et al. (31) utilized recurrent neural networks and historical TC information (HIS) to forecast the RI TCs. However, their forecasts were less effective due to excluding environmental factors. Cloud et al. (32) employed the multilayer perceptrons method using the SHIPS model’s forecast factors as inputs, improving the RI TC forecasting performance in the Atlantic and East Pacific Oceans. Yang et al. (33) utilized long short-term memory (LSTM) networks (34) to construct the RI TC forecasting model based on the SHIPS (23) model’s forecast factors. Su et al. (19) and Griffin et al. (5) considered satellite infrared images in addition to the SHIPS model’s forecast factors, enhancing the forecasting performance of RI TC periods in the Atlantic and East Pacific. Chen et al. (20) developed a DL-based ensemble forecasting method that demonstrated promising potential to improve forecast accuracy. However, this model was tested in a controlled environment and did not utilize real-time data inputs in operational settings. Zhou et al. (21) proposed a fusion model based on satellite infrared images of CNN and LSTM, improving the RI TC periods forecasting performance in the Northwest Pacific Ocean.
Common metrics used to evaluate the performance of RI TC forecasts include POD, FARate (False Alarm Rate), and FARatio (false alarm ratio). POD is the proportion of actual events that were correctly forecasted. FARate is the proportion of false alarms among all negative forecasts. FARatio is the proportion of forecasted events that did not occur. These three metrics will be used to evaluate the RI TC forecasting capability of various methods.
Compared to the numerical and statistical models, the DL models based on satellite infrared and the SHIPS model’s forecast factors greatly enhance the RI TC POD, with the best RI TC POD reaching 82.6% (21). However, the highest performance of the above-mentioned DL-based methods had a significantly high FARate (27.2%) for non-RI TC periods (21). First, RI TC only accounts for about 5% of the total TC samples, leading to an imbalance between RI TC and non-RI TC data (35). The imbalance in data quantities can result in biased model performance, e.g., the high non-RI TC FARate. This issue is still outstanding as of now (5, 19, 21). Second, The TC intensity change is highly sensitive to environmental factors; however, existing DL methods only use the maximum and mean values of environmental factors, failing to represent the spatial structure of a TC.
Here, we developed a contrastive-based RI TC forecasting model named the RITCF-contrastive model using satellite infrared imagery (SAT), atmospheric, and oceanic data. The model solves sample imbalance through contrastive learning (36), significantly enhances the POD, and reduces the FARate for RI TC forecasting. Second, using three-dimensional (3D) environmental factors allows the model to extract TC spatial features.
The RITCF-contrastive model was developed using satellite infrared, atmospheric, and oceanic data from 2000 to 2017 (303 RI TC periods and 6492 non-RI TC periods), validated with data from 2018 to 2019 (110 RI TC periods and 1,362 non-RI TC periods), and tested using data from 2020 to 2021 (78 RI TC periods and 1,071 non-RI TC periods). Training data update the model parameters, validation data help prevent overfitting during training, and test data are used for the final evaluation. Dividing the entire dataset into training, validation, and testing segments is a standard procedure in developing deep-learning models. While we have not validated the model using real-time data, our analysis of the results based on the reanalysis data and near-real-time operational data indicates that the model demonstrates considerable potential. The model produces reliable outcomes, achieving a RI TC POD of 92.3%, a non-RI TC FARate of 8.9%, and a non-RI TC FARatio of 56.9%.
Results
Forecasting Results of RI TC.
The RITCF-contrastive model consists of two parts of input, named Input-A and Input-B. Input-A represents a known RI TC sample, while Input-B represents the forecasted unknown sample. The RITCF-contrastive model forecasts RI TC by analyzing the similarities and differences between Input-A and Input-B. For example, if the characteristics of Input-A and Input-B are similar, the RITCF-contrastive model forecasts that they belong to the same category. To enhance the model’s accuracy, each sample in the test data was regarded as an unknown and paired with 10 different known RI TC periods, forming 10 different sample pairs. These pairs were then inputted into the RITCF-contrastive model, yielding 10 predictions for each unknown sample. The process was referred to as voting. If six or more out of the 10 predictions classify the sample as RI TC, it is forecasted as RI TC; otherwise, it is classified as non-RI TC.
Fig. 1 A and B show the RITCF-contrastive model test results. The inputs include the u- and v-component of winds (U and V), potential vorticity (PV), SST, SAT, and HIS forecast factors. Fig. 1A shows the forecasting results for RI TC periods. 71 RI TC periods received six or more votes and were correctly forecasted as RI TC periods. The RITCF-contrastive model achieves a high RI TC POD (91.0%). The seven misclassified RI TC periods were from four TC cases. As shown in Fig. 1A, six RI TC periods obtained five votes and were incorrectly forecasted as non-RI TC periods by one vote, 1 RI TC period received four votes and was also incorrectly forecasted as a non-RI TC. The misforecast RI TC periods are only missed because of the difference of one or two votes. The results indicate that the RITCF-contrastive model can accurately identify the features of RI TC periods.
Fig. 1.
Fig. 1B shows the forecasting results for non-RI TC periods. The 2020–2021 Northwest Pacific test data included 1,071 non-RI TC periods. Further, 970 non-RI TC periods were correctly forecasted, while 101 non-RI TC periods were incorrectly forecasted. The RITCF-contrastive model achieves a low non-RI TC FARate (9.4%) and a low non-RI TC FARatio (58.7%). Unlike the RI TC periods voting pattern, the non-RI TC periods votes are concentrated at 0 and 10 votes. It indicates that the RITCF-contrastive model has strong discriminatory performance for most non-RI TC periods and RI TC periods. However, some non-RI TC periods are consistently misforecast for specific reasons, which will be discussed in the following section on the analysis of misforecasted non-RI and RI TC periods.
Fig. 1C shows the locations of RI TC occurrences and the forecast results from the RITCF-contrastive model during 2020–2021. The occurrence of RI TC periods is mainly concentrated in the regions west of 140°E and south of 24°N. The blue dots indicate the locations where RI TC periods were accurately forecasted, while the red dots show where the forecasts were incorrect. The incorrectly forecasted RI TC periods are primarily located between 125 to 150°E and 16 to 20°N. The RITCF-contrastive model effectively identifies most RI TC periods near coastal areas for the test dataset, offering crucial early warnings that aid in disaster prevention and mitigation.
Comparison of RITCF-Contrastive Model with other Methods.
Table 1 highlights the advantage of the RITCF-contrastive model in forecasting RI TC periods compared to traditional and DL-based methods. DL-based methods (19–21) demonstrate greater potential compared to traditional methods in improving RI TC POD. However, the non-RI TC FARate and FARatio are higher than those of traditional methods due to the imbalanced RI TC and non-RI TC periods. Our RITCF-contrastive model, based on contrastive learning, greatly improves the issue of data imbalance by constructing sample pairs. The RITCF-contrastive model obtains an RI TC POD of 91.0% a non-RI TC FARate of 9.4% and a non-RI TC FARatio of 58.7%. Compared to other DL-based methods, improving RI TC POD reduces non-RI TC FARate by more than half.
Table 1.
| Method | Test data | RI TC POD, % | RI TC FA ratio, % | non-RI TC FA rate, % | |
| DL | Su et al. (19) | 2009-2014 North Atlantic/Eastern Pacific | 65/85 | 77/65 | \ |
| Zhou et al. (21) | 2019-2021 Northwest Pacific | 82.6 | 75 | 27.2 | |
| Chen et al. (20) | 2017 Northwest Pacific | 37.5 | \ | 6.7 | |
| RITCF-contrastive (input U, V, PV, SST, SAT, HIS) | 2020-2021 Northwest Pacific | 91.0 | 58.7 | 9.4 | |
| RITCF-contrastive-operational (input U, V, PV, SST, SAT, HIS) Use IFS-Operational data (IFS oper 0:00 and 12:00 at step = 0 h, scda 6:00 and 18:00 at step = 0 h) | 2020-2021 Northwest Pacific | 92.3 | 57.9 | 9.2 | |
| RITCF-contrastive-operational (input U, V, PV, SST, SAT, HIS) Use IFS-Operational data (IFS oper 0:00 and 12:00 at step = 0 and 6 h) | 2020-2021 Northwest Pacific | 92.3 | 56.9 | 8.9 | |
| Traditional | CMA (21) | 2019 Northwest Pacific | 51.6 | 83.2 | 8.1 |
| NOAA/NCEP (21) | 43.1 | 70.3 | 9.9 |
CMA means China Meteorological Administration, NOAA/NCEP means National Oceanic and Atmospheric Administration/National Centers for Environmental Prediction.
The results from Chen et al. (20), Zhou et al. (21), and the RITCF-contrastive model used data (ERA5 and TC best track data) that are not available in real-time, so they don’t reflect the method’s performance in actual operational forecasting. In contrast, Su et al. (19) used data more representative of real-time forecasting. To better simulate the operational forecasting performance of the RITCF-contrastive model, we replaced ERA5 reanalysis data and CMA (China Meteorological Administration) best track data with operationally available data as input (see the section The RITCF-Contrastive-Operational Model for details).
Surprisingly, with IFS-operational data, the model achieved a higher RI TC POD (92.3%) and lower non-RI TC FARate (8.9%) and FARatio (56.9%). While both ERA5 and IFS-Operational incorporate data assimilation, there are significant differences. For example, ERA5 ingests additional late-arriving observations not available in time for the operational IFS. Furthermore, ERA5 uses an older version of the IFS model with coarser resolution. These factors may lead to an underestimation of maximum wind speed in ERA5, resulting in lower wind speeds compared to IFS-Operational data (SI Appendix, Fig. S1).
Since our model forecasts RI TC by comparing Input-A and Input-B, ERA5’s underestimation of maximum wind speed can lead to misjudgments in intensity forecasting. In contrast, IFS-Operational data, with less smoothing, aligns more closely with the actual maximum wind speed of the TC at the time of satellite imagery. The result leads to more pronounced differences between input-A and input-B in IFS-Operational data, reducing the model’s error in forecasting RI cases and explaining the model’s improved performance with IFS-Operational data compared to ERA5.
The Misforecasted RI TCs.
The misforecasted RI TC periods were analyzed through Input-A/Input-B. Fig. 2 compares the change in 24 h (C24) of Input-A/Input-B for RI TC periods that were correctly forecasted and misforecasted. For example, the difference between Input-A/Input-B at t = 0 h and t = −24 h was calculated first to determine C24. Then, the C24 across test data samples was averaged to generate C24 for Input-A (C24A) and C24 for Input-B (C24B) in Fig. 2. Fig. 2A shows C24A and C24B for misforecasted RI TC periods in test data. One can see that C24A and C24B, especially in U, V, and SST factors, are very different. Fig. 2B compares C24B for correctly forecasted with C24B for misforecasted RI TC periods, revealing significant differences. As can be seen from wind and pressure’s C24B, RI TC periods with low TC intensity (red line in Fig. 2B) are prone to misforecasting. It means that forecasters need to be careful when using RITCF-contrastive model to identify TC periods that have varied in intensity between 15 and 23 m/s over the past 24 h, as the model may generate false alarms. Fig. 2C contrasts C24A and C24B for correctly forecasted RI TC periods highlighting their similarities. The RITCF-contrastive model accurately identifies RI TC features, correctly forecasting over 90% of RI TC periods. The data show that non-RI TC periods with intensity changes near the RI threshold frequently lead to misforecasts. The model’s reliance on contrasting similarities between Input-A and Input-B to forecast RI TC periods complicates efforts to lower the FARate and FARatio for non-RI TC periods, remaining a significant challenge.
Fig. 2.
The Misforecasted Non-RI TCs.
Fig. 3 analyzes misforecasted non-RI TC periods, similar to Fig. 2, by comparing C24A/C24B for correctly forecasted and misforecasted non-RI TC periods. Fig. 3A calculated the C24A and C24B for misforecasted non-RI TC periods in the test data. These images reveal that misforecast non-RI TC periods are in an intensification phase with similar Input-A and Input-B inputs. Fig. 3B highlights the C24B for correctly forecasted and misforecasted non-RI TC periods in test data, showing notable distinctions. Compared to correctly forecasted non-RI TC periods, the misforecasted non-RI TC periods were in a wind-intensifying (wind’s C24B in Fig. 3B), small-SST cooling (SST’s C24B in Fig. 3B) and positive brightness temperature change (SAT’s C24B in Fig. 3B) phase. When encountering wind-intensifying, small-SST cooling, and positive brightness temperature change, forecasters can apply their expertise to adjust the model’s forecast results, thereby reducing the FARate and FARatio. Fig. 3C displays the C24A and C24B for correctly forecasted non-RI TC periods in test data. Evidently, these inputs are significantly different, leading to accurately identified by the RITCF-contrastive model. In conclusion, the similarity between RI TC periods and misforecasted non-RI TC periods leads to inaccuracies in the RITCF-contrastive model’s forecasts.
Fig. 3.
Discussion
Contrastive learning effectively addresses the data imbalance challenge in RI TC forecasting by constructing balanced sample pairs, avoiding redundancy and bias. The RITCF-contrastive model reduces the FARate and FARatio for non-RI TC cases while maintaining a high detection rate for RI TCs, thereby enhancing forecast precision by simplifying the differentiation between RI and non-RI cases.
Sample imbalance is a significant limitation in DL-based RI TC forecasting, as the scarcity of RI TC samples leads to models favoring the more common non-RI cases, hindering their ability to identify critical RI TC predictors. Previous studies (e.g., Su et al., Zhou et al., Chen et al.) have suggested methods such as reducing non-RI samples, duplicating RI samples, or modifying the loss function to emphasize RI TCs. However, these approaches may cause underfitting, distort data distribution, or introduce biases that increase the FARate and FARatio for non-RI cases. In contrast, the RITCF-contrastive model employs contrastive learning to create balanced pairs of RI and non-RI samples, enabling the model to learn distinct spatial representations and improve predictive accuracy. By mapping samples to one-dimensional vectors, it assesses feature similarities or differences, simplifying the forecasting task and enhancing precision.
A key limitation of traditional RI TC forecasting methods is their inability to adequately represent TC structural features. SHIPS-based methods reduce environmental factors to one-dimensional features, overlooking crucial spatial information about the storm’s structure. The RITCF-contrastive model addresses this by utilizing 3D atmospheric and oceanic data to capture both spatial and temporal characteristics of TCs, resulting in more accurate forecasts. Fig. 4 illustrates the roles of dynamical, thermodynamical, and structural factors in RI TC forecasting, which aligns with existing knowledge. Compared to previous DL models, the RITCF-contrastive model maintains a high POD for RI TCs while significantly reducing the FARate and FARatio for non-RI cases, demonstrating its robustness even in data-limited situations.
Fig. 4.
The RITCF-contrastive model has several limitations: 1) TC characteristics vary significantly across ocean basins, as noted by the Joint Typhoon Warning Center, which divides the Pacific into regions with separate models. This study only validated the model’s performance in the Northwest Pacific between 2020 and 2021. It would be worth mentioning that testing a larger sample and more ocean basins is needed to test the RITCF-contrastive model. 2) The model relies on complex multimodal forecasting factors, posing challenges for interpretability. While current AI interpretability methods struggle to clarify model results, this does not detract from the RITCF-contrastive model’s scientific value. As noted by Homl et al. (37) and Ham et al. (38), AI can lead to discoveries that challenge or advance existing theories. The RITCF-contrastive model offers a different approach for enhancing RI TC forecasting accuracy, with promising potential for practical applications.
Materials and Methods
Data.
The RI TC forecast model (RITCF-contrastive) inputs include satellite infrared images, atmospheric and oceanic environmental factors, and TC historical information. The satellite infrared images are from GridSat-B1 (the Gridded Satellite B1), the atmospheric and oceanic environmental factors are from the ERA5 dataset (the fifth generation of reanalysis data from the European Centre for Medium-Range Weather Forecasts), and the TC historical information is from China Meteorological Administration (CMA) best track data.
The GridSat-B1 dataset (39, 40) originates from the International Satellite Cloud Climatology Project, aiming to facilitate the use of satellite observations of the Earth. The GridSat-B1 dataset is widely used in TC research (28). The GridSat-B1 dataset consists of composite images from multiple geostationary satellites (including SMS-2, GOES-1 to GOES-15, Meteosat-2 to Meteosat-10, GMS-1 to GMS-5, MTSAT-1R, MTS-2, FY2-C/E). If a latitude and longitude grid point has observations from multiple satellites simultaneously, the lowest brightness temperature value among the observations is used to generate the final image, resulting in global coverage of infrared imagery. The GridSat-B1 dataset provides global infrared brightness temperature data every 3 h with a spatial resolution of 0.07° from 1980 to the present, including infrared (~11 μm), visible (~0.6 μm), and water vapor (~6.7 μm) bands.
The second dataset is the ERA5 dataset (41–43), covering atmospheric and oceanic reanalysis data from 1940 to the present. The ERA5 dataset has a temporal resolution of 1 h and a spatial resolution of 0.25°. We downsampled the spatial resolution to 1° and selected data points at 00, 06, 12, and 18 UTC daily, resulting in a temporal resolution of 6 h. The atmospheric reanalysis data consist of 37 vertical levels. Factors that influence TC intensity changes are observed across different levels of the atmospheric vertical profile, with 1,000, 850, 500, and 200 hPa representing the surface, lower, middle, and upper atmosphere. These levels effectively reflect the 3D structure and dynamic processes of TC periods, which can help forecast the evolution of TC intensity (28). Therefore, the atmosphere factors at 1,000, 850, 500, and 200 hPa were selected for RI TC forecasting research. The atmospheric reanalysis data selected for the paper include u-component of winds (U), v-component of winds (V), vertical velocity (W), temperature (T), PV, divergence (D), geopotential (G), and relative humidity (RH). The sea surface reanalysis data include SST.
The third dataset is the CMA best track dataset (44, 45), provided by the CMA’s Tropical Cyclone Data Center and compiled by the Shanghai Typhoon Institute, adhering to the standards outlined in the “Regulations on Typhoon Operations and Services”. The CMA best track dataset provides information on Northwest Pacific TC periods from 1949 to the present, including the TC’s center, 2 min maximum sustained wind speed, minimum sustained pressure, etc.
The fourth dataset is the CMA operational data (https://typhoon.weather.com.cn/) (46). We obtain the CMA operational data on Northwest Pacific TC periods from 2020 to 2021, including the TC’s center, 2 min maximum sustained wind speed, minimum sustained pressure, etc.
The fifth dataset is the IFS-Operational (ECMWF Integrated Forecast System, Operational version) forecast data (https://apps.ecmwf.int/archive-catalogue/?class=od&stream=oper and https://apps.ecmwf.int/archive-catalogue/?class=od&stream=scda) (47, 48), including oper (Operational Forecast) and scda (Single-Column Data Assimilation) experiment. The atmospheric and oceanic forecast factors were same as ERA5 dataset.
In the study, the 95th percentile of 24 h TC intensity change serves as the threshold, defining RI TC periods in the Northwest Pacific as those with a 24 h intensity change greater than or equal to 13 m/s (accounting for 5.4% of TC periods from 2000 to 2021).
Data Processing.
As shown in Table 2, U, V, W, geopotential (G), PV, divergence (D), temperature (T), RH, SST, SAT, and HIS were considered as candidate forecasting factors. Forecasting factors from the ERA5 dataset were extracted from four isobaric levels within a 25° × 25° area centered around the TC’s center at time points t = 0, −6, −12, −18, and −24. The SAT data were sourced from the GridSat-B1 dataset, extracted for time points t = 0, −6, −12, −18, and −24, in images sized 224 × 224 pixels (about 0.07° spatial resolution) centered around the TC’s center. HIS is from the CMA best track dataset, including the maximum sustained wind speed and minimum sustained pressure at time points t = 0, −6, −12, −18, and −24, along with their differences.
Table 2.
| Source | Time | Isobaric level | |
|---|---|---|---|
| U (u-component of winds) | ERA5 | t = 0, −6, −12, −18, −24 | z = 200, 500, 850, 1,000 hpa |
| V (v-component of winds) | ERA5 | ||
| W (w-component of winds) | ERA5 | ||
| G (geopotential) | ERA5 | ||
| PV | ERA5 | ||
| D (divergence) | ERA5 | ||
| T (temperature) | ERA5 | ||
| RH (relative humidity) | ERA5 | ||
| SST | ERA5 | / | |
| SAT (satellite infrared image) | GridSat-B1 | ||
| HIS | CMA best track | wind0, wind6, wind12, wind18, wind24, pres0, pres6, pres12, pres18, pres24, wind0-wind6, wind0-wind12, wind0-wind18, wind0-wind24, wind6-wind12, wind6-wind18, wind6-wind24, wind12-wind18, wind12-wind24, wind18-wind24, pres0-pres6, pres0-pres12, pres0-pres18, pres0-pres24, pres6-pres12, pres6-pres18, pres6-pres24, pres12-pres18, pres12-pres18, pres18-pres24, | |
Comparative learning is used for RI TC forecasting, which necessitates the construction of sample pairs. In a sample pair, if both the Input-A input and the Input-B input are RI TC periods, the label is “0”. If the Input-A input is an RI TC period and the Input-B input is a non-RI TC period, the label is “1”.
The construction of sample pairs follows the steps: Taking the training data as an example (training data has 303 RI TC periods and 6,492 non-RI TC periods), RI TC period A is selected along with the 21 closest RI TC periods in time, forming 21 positive sample pairs, resulting in a total of 303 × 21 = 6,363 positive sample pairs. For example, We arrange all RI cases in the training data in chronological order to create the RI-cases-time list [e.g., (1, 2, 3, 4, 5, 6, 7, 8, …, 100, 101, 102, …)], where each index corresponds to an RI TC period. For example, if Input-B is an RI case on October 1, 2002, and its index in the RI-case-time list is 30, then we select 21 RI cases for Input-A from those indexed before 30, and that occurred no later than September 30, 2002. These samples are typically from different TC cases because most TCs have fewer than 7 RI cases. Non-RI TC period is selected along with the 1 closest RI TC period in time, forming 1 negative sample pair, resulting in a total of 6,492 × 1 = 6,492 negative sample pairs. The construction of validation data follows the same procedure as the training data, resulting in 1,320 positive and 1,362 negative sample pairs. The 1,149 TC periods from 2020 and 2021 are considered unknown samples for the test data. Sample pairs are formed by matching them with the 10 closest RI TC periods in time, and the unknown case categories are determined through voting. Finally, 12,855 training data were obtained (6,363 positive sample pairs and 6,492 negative sample pairs), along with 2,682 validation data (1,320 positive sample pairs and 1,362 negative sample pairs), and 11,490 test data (780 positive sample pairs and 10,710 negative sample pairs).
We need to follow the following principles when constructing sample pairs: For a given sample of Input-B (unknown case), Input-A (known case) does not include cases later in time than Input-A. A known case means that at the time of forecasting, whether it is an RI or non-RI TC is already known. Since determining whether a TC undergoes RI requires knowing the change in TC intensity over the next 24 h, we impose a strict condition: The samples of Input-A and Input-B must be separated by at least 24 h. For example, suppose we are forecasting whether a TC will rapidly intensify at 00:00 on October 1, 2019. In that case, the Input-A sample must have occurred no later than 00:00 on September 30, 2019, to ensure that the sample is known and no future data are used.
Normalization of data is imperative before training neural networks for accelerated computations and achieving desirable outcomes (28). In the paper, input data underwent a linear transformation to conform to the range of [0, 1] for T, PV, D, RH, SST, and SAT; [−1, 1] for U, V, and W (directional variables) through the following function:
[1]
[2]
where xmin and xmax represent the minimum and maximum values of U, V, W, SST, or IR, and y denotes the normalized value, constrained within the range of [−1, 1]. The wind components U, V, and W are normalized to a range of −1 to 1 to reflect their directional attributes, with positive values indicating east, north, and upward directions, and negative values indicating the opposite. In contrast, other variables without directional attributes are normalized to a range of 0 to 1. This dual normalization approach ensures that both ranges fall within the activation range of DL functions, facilitating model training. This strategy was chosen to maintain the integrity of directional data while optimizing the overall input format.
The RITCF-Contrastive Model.
RITCF-contrastive model structure, based on contrastive learning (35) with 3D environmental factors input, is shown in Fig. 5. The input consists of a pair of samples (Input-A and Input-B). The RITCF-contrastive model is a DL method that realizes classification tasks by comparing the similarities and differences between samples. Its core idea is to bring similar samples or features closer to each other in the feature space while contrasting different samples or features to push them further apart in the feature space. The “feature space” in DL is like an image map where each point represents an item’s characteristics, helping to visualize and organize data by similarity. The number of sample pairs can be greatly increased by constructing pairs of two different samples, thereby addressing the issue of sample imbalance. All inputs include data from time points t = −24, −18, −12, −6, 0 h, and variables from isobaric levels 200, 500, 850, and 1,000 hPa: U, V, PV, SST, SAT, and HIS. The RITCF-contrastive of module Input-A and Input-B use convolutional, pooling, fully connected layers, add, and concatenate to map the two inputs to an n × 1 vector. Then, the RITCF-contrastive model computes the spatial distance between these two vectors. For example, as shown in Fig. 5, Parts A and B outputs are 32×1 vectors x and y. Part C calculates the spatial distance between x and y to yield the model output, ranging between 0 and 1. If the output is close to 0, it indicates that the two samples in the sample pair are in the same category. Conversely, if the output is close to 1, it indicates that the two samples in the sample pair have been mapped to distant regions and are of different categories.
Fig. 5.
Compared to existing RI TC forecast methods (Table 1), our study presents two key points of innovation: First, the contrastive learning method has been employed to address data imbalance and enhance the forecast accuracy of RI TC periods; Second, the structure of RITCF-contrastive can better extract spatiotemporal correlation features from multimodal data (3D environmental factors, 2D satellite images, and 1D HIS) and achieve better forecasting results.
Model Evaluation Metrics.
POD, false alarm rate (FARate), and false alarm ratio (FARatio) were used to evaluate the performance of RI TC forecasting methods. If we defined a = number of events forecasted and observed; b = number of events forecasted but not observed; c = number of events not forecasted but observed; d = number of events forecasted and not observed. Then, POD = a/(a + c); FARate = b/(b + d); FARatio = b/(a + b).
Steps of Experiment.
The selection of forecast factors can influence model performance. We design sensitivity experiments to explore the roles of different forecasting factors in the RI TC forecast and determine the best factors for the RITCF-contrastive model.
RI TC forecasts in the testing data are conducted through a voting mechanism. The RITCF-contrastive model regards 1,149 TC periods as unknown samples. Then, the model pairs each unknown sample with the 10 nearest known RI TC periods in time, forming sample pairs. Each unknown sample is input into the RITCF-contrastive model 10 times to generate 10 voting results, determining whether it is forecasted as RI or non-RI TC.
Tensorflow and Keras were used to build and train the RITCF-contrastive model. The RITCF-contrastive model uses the linear activation function in the output layer and ReLU for the other layers. The optimization function selected is Adam, while the loss function employed is Mean Squared Error.
Sensitivity experiments analyzed basic forecast factors like U, V (TC dynamic factors), SST (thermal factor), HIS, and SAT (structural factors). Additional inputs include W, D, PV, G, RH, and T. The RITCF-contrastive-1 model, incorporating the W factor, achieves an RI TC POD of 93.6% and a non-RI TC FARate of 12.1%. The RITCF-contrastive-2 model improves RI TC POD to 98.7% but increases non-RI TC FARate to 12.7%. Models RITCF-contrastive-3, -4, -5, and -6 show a slight decrease in RI TC POD (about 4%, resulting in 2 to 3 additional missed RI TC periods) but significantly decrease the non-RI TC FARate (about 2.4%, reducing 25 false alarms). RITCF-contrastive-7 and -8, adding G and T to U, V, SST, SAT, HIS, and PV, result in RI TC POD decreases to 85.9% and 84.6%, and non-RI TC FARate increases to 13.4% and 14.3%. RITCF-contrastive-9, replacing U and V with G and T, sees a slight RI TC POD decrease (84.6%) and a significant non-RI TC FARate increase (17.2%). This performance drop is due to intercorrelation among forecast factors and increased model complexity, which leads to overfitting. The results show that D improves RI TC POD but increases non-RI TC FARate, RH improves RI TC POD and non-RI TC FARate, and PV, G, and T reduce RI TC POD but significantly decrease non-TC FARate. The optimal combination is balancing RI TC POD and non-RI TC FARate, RITCF-contrastive-3.
The Operational Version RITCF-Contrastive Model.
The RITCF-contrastive model was tested by using IFS-Operational data and CMA operational data. It is important to note that, although the evaluation was conducted using operational data, the model itself (including all model parameters) remained unchanged and was still trained on ERA5 reanalysis data. In the operational version RITCF-contrastive model: (1) The TC intensity was sourced from CMA operational data instead of CMA best track data. (2) IFS-Operational forecast data were used to replace ERA5 data. Since IFS-Operational data have two versions, the IFS-Operational oper experiment provides 10-d forecast results at 00:00 and 12:00 daily (steps from 0 to 240 h), and the IFS-Operational scda experiment provides 4-d forecast results at 06:00 and 18:00 daily (steps from 0 to 90 h). Two substitution schemes were used:
Exercise 1: Replacing ERA5 reanalysis data with step=0 h IFS-Operational data. For example, the ERA5 data at 00:00 on January 1, 2019, was replaced by the step=0 h IFS-Operational oper forecast at 00:00, and the ERA5 data at 06:00 was replaced by the step=0 h IFS-Operational scda forecast at 06:00.
Exercise 2: Replacing ERA5 reanalysis data with step=0 h and step=6 h IFS-Operational data. For instance, the data at 00:00 on January 1, 2019, was replaced by the step=0 h IFS-Operational oper forecast at 00:00, and the data at 06:00 was replaced by the step=6 h IFS-Operational oper forecast at 00:00.
Data, Materials, and Software Availability
The GridSat-B1 dataset is from www.ncei.noaa.gov/products/gridded-geostationary-brightness-temperature (40). The ERA5 reanalysis data are available from https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=form (42) and https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-pressure-levels?tab=form (43). The CMA best track data are from https://tcdata.typhoon.org.cn/zjljsjj.html (45). The CMA operational data are available from https://typhoon.weather.com.cn/ (46). The IFS-Operational forecast data are available from https://apps.ecmwf.int/archive-catalogue/?class=od&stream=oper (47) and https://apps.ecmwf.int/archive-catalogue/?class=od&stream=scda (48).
Acknowledgments
This work was supported by the NSFC Innovative Group Grant (No. 42221005), the National Natural Science Foundation of China (grant nos. U2006211, 42376175, 42090044, and 42076200), the Strategic Priority Research Program of the Chinese Academy of Sciences (grant no. XDB42040401), the China-Portugal Xinghai “Belt and Road” Joint laboratory and joint research on new air and sea technologies (2022YFE0204600), and the Postdoctoral Fellowship Program of CPSF (GZC20241741).
Author contributions
C.W. and X.L. designed research; C.W. performed research; C.W. and N.Y. analyzed data; and C.W. wrote the paper.
Competing interests
The authors declare no competing interest.
Supporting Information
Appendix 01 (PDF)
- Download
- 288.86 KB
References
1
K. Balaguru, G. R. Foltz, L. R. Leung, Increasing magnitude of hurricane rapid intensification in the central and eastern tropical Atlantic. Geophys. Res. Lett. 45, 4238–4247 (2018).
2
K. T. Bhatia et al., Recent increases in tropical cyclone intensification rates. Nat. Commun. 10, 1–9 (2019).
3
Y. Li et al., Recent increases in tropical cyclone rapid intensification events in global offshore regions. Nat. Commun. 14, 5167 (2023).
4
U. Shimada et al., Further improvements to the Statistical Hurricane Intensity Prediction Scheme using tropical cyclone rainfall and structural features. Weather Forecasting 33, 1587–1603 (2018).
5
S. M. Griffin, A. Wimmers, C. S. Velden, Predicting rapid intensification in North Atlantic and eastern North Pacific tropical cyclones using a convolutional neural network. Weather Forecasting 37, 1333–1355 (2022).
6
J. Xu, Y. Wang, Dependence of tropical cyclone intensification rate on sea surface temperature, storm intensity, and size in the western North Pacific. Weather Forecasting 33, 523–537 (2018).
7
J. Kaplan et al., Evaluating environmental impacts on tropical cyclone rapid intensification predictability utilizing statistical models. Weather Forecasting 30, 1374–1396 (2015).
8
E. A. Hendricks, Internal dynamical control on tropical cyclone intensity variability–a review. Tropical Cyclone Res. Rev. 1, 97–105 (2012).
9
M. DeMaria, A simplified dynamical system for tropical cyclone intensity prediction. Mon. Weather Rev. 137, 68–82 (2009).
10
M. DeMaria, C. R. Sampson, J. A. Knaff, K. D. Musgrave, Is tropical cyclone intensity guidance improving? Bulletin of the Am. Meteorol. Soc. 95, 387–398 (2014).
11
J. P. Cangialosi et al., Recent progress in tropical cyclone intensity forecasting at the National Hurricane Center. Weather Forecasting 35, 1913–1922 (2020).
12
M. S. Fischer, B. H. Tang, K. L. Corbosiero, C. M. Rozoff, Normalized convective characteristics of tropical cyclone rapid intensification events in the North Atlantic and eastern North Pacific. Mon. Weather Rev. 146, 1133–1155 (2018).
13
E. N. Rappaport, J.-G. Jiing, C. W. Landsea, S. T. Murillo, J. L. Franklin, The Joint Hurricane Test Bed: Its first decade of tropical cyclone research-to-operations activities reviewed. Bull Am. Meteorol. Soc. 93, 371–380 (2012).
14
H. Wang, Y. Wang A numerical study of Typhoon Megi (2010). Part I: Rapid intensification. Mon. Weather Rev. 142, 29–48 (2014).
15
V. Tallapragada, “Overview of the NOAA/NCEP operational hurricane weather research and forecast (HWRF) modelling system” in Advanced Numerical Modeling and Data Assimilation Techniques for Tropical Cyclone Prediction, U. C. Mohanty, S. G. Gopalakrishnan, Eds. (Springer, Dordrecht, 2016), chap. 51-106.
16
J. Kaplan, M. DeMaria, J. A. Knaff, A revised tropical cyclone rapid intensification index for the Atlantic and eastern North Pacific basins. Weather Forecasting 25, 220–241 (2010).
17
J. Kaplan, M. DeMaria, Large-scale characteristics of rapidly intensifying tropical cyclones in the North Atlantic basin. Weather Forecasting 18, 1093–1108 (2003).
18
E. A. Hendricks, M. S. Peng, B. Fu, T. Li, Quantifying environmental control on tropical cyclone intensity change. Mon. Weather Rev. 138, 3243–3271 (2010).
19
H. Su et al., Applying satellite observations of tropical cyclone internal structures to rapid intensification forecast with machine learning. Geophys. Res. Lett. 47, e2020GL089102 (2020).
20
B. F. Chen, Y. T. Kuo, T. S. Huang, A deep learning ensemble approach for predicting tropical cyclone rapid intensification. Atmospheric Sci. Lett. 24, e1151 (2023).
21
G. Zhou, J. Xu, Q. Qian, Y. Xu, Y. Xu, Discriminating technique of typhoon rapid intensification trend based on artificial intelligence. Atmosphere 13, 448 (2022).
22
S. Gopalakishnan et al. 2020 HFIP R&D Activities Summary: Recent Results and Operational Implementation (2021). https://doi.org/10.25923/718e-6232
23
M. DeMaria, M. Mainelli, L. K. Shay, J. A. Knaff, J. Kaplan, Further improvements to the statistical hurricane intensity prediction scheme (SHIPS). Weather Forecasting 20, 531–543 (2005).
24
X. Li et al., Deep-learning-based information mining from ocean remote-sensing imagery. Natl. Sci. Rev. 7, 1584–1605 (2020).
25
G. Zheng, X. Li, R. H. Zhang, B. Liu, Purely satellite data–driven deep learning forecast of complicated tropical instability waves. Sci. Adv. 6, eaba1482 (2020).
26
L. Zhou, R.-H. Zhang, A self-attention–based neural network for three-dimensional multivariate modeling and its skillful ENSO predictions. Sci. Adv. 9, eadf2827 (2023).
27
H. Wang, X. Li, DeepBlue: Advanced convolutional neural network applications for ocean remote sensing. IEEE Geosci. Remote Sens. Mag. 12, 138–161 (2023).
28
C. Wang, X. Li, G. Zheng, Tropical cyclone intensity forecasting using model knowledge guided deep learning model. Environ. Res. Lett. 19, 024006 (2024).
29
C. Wang, Q. Xu, Y. Cheng, Y. Pan, H. Li, Ensemble forecast of tropical cyclone tracks based on deep neural networks. Front Earth Sci. 16, 671–677 (2022).
30
Y. Li et al., Leveraging LSTM for rapid intensifications prediction of tropical cyclones. ISPRS Ann. Photogram., Remote Sens. Spatial Inform. Sci. 4, 101–105 (2017).
31
R. Chandra, Towards prediction of rapid intensification in tropical cyclones with recurrent neural networks. arXiv [Preprint] (2017). https://api.semanticscholar.org/CorpusID:19024188 (Accessed 2 November 2022).
32
K. A. Cloud et al., A feed forward neural network based on model output statistics for short-term hurricane intensity prediction. Weather Forecasting 34, 985–997 (2019).
33
Q. Yang, C.-Y. Lee, M. K. Tippett, A long short-term memory model for global rapid intensification prediction. Weather Forecasting 35, 1203–1220 (2020).
34
A. Graves, “Long Short-Term Memory” in Supervised Sequence Labelling with Recurrent Neural Networks: Studies in Computational Intelligence (Springer, Berlin, 2012), chap. 37-45.
35
T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollár, “Focal loss for dense object detection” in Proceedings of the IEEE International Conference on Computer Vision, (2017) pp 2980-2988.
36
A. Jaiswal, A. R. Babu, M. Z. Zadeh, D. Banerjee, F. Makedon, A survey on contrastive self-supervised learning. Technologies 9, 2 (2020).
37
E. A. Holm, In defense of the black box. Science 364, 26–27 (2019).
38
Y.-G. Ham, J.-H. Kim, J.-J. Luo, Deep learning for multi-year ENSO forecasts. Nature 573, 568–572 (2019).
39
C. F. Dickinson, C. N. Helms, C. C. Hennon, C. D. Holmes, Globally gridded satellite (GridSat) observations for climate studies 2. BAMS 92, 893–907 (2011).
40
K. R. Knapp et al., Data from “Globally gridded satellite (GridSat) observations for climate studies.” NOAA. https://www.ncei.noaa.gov/products/gridded-geostationary-brightness-temperature. Deposited 4 April 2023.
41
Copernicus Climate Change Service, ERA5 Hourly Data on Pressure Levels from 1940 to Present. (CDS, 2023).
42
H. Hersbach et al., Data from “ERA5 hourly data on single levels from 1940 to present.” Copernicus Climate Change Service (C3S) Climate Data Store (CDS). https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=form. Deposited 4 April 2023.
43
H. Hersbach et al., Data from “ERA5 hourly data on pressure levels from 1940 to present.” Copernicus Climate Change Service (C3S) Climate Data Store (CDS). https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-pressure-levels?tab=form. Deposited 4 April 2023.
44
M. Ying et al., An overview of the China meteorological administration tropical cyclone database. J. Atmospheric and Oceanic Technol. 31, 287–301 (2014).
45
X. Lu et al., Data from “Western North Pacific tropical cyclone database created by the China Meteorological Administration.” CMA Tropical Cyclone Data Center. https://tcdata.typhoon.org.cn/zjljsjj.html. Deposited 4 April 2023.
46
CMA, Data from: “Typhoon monitoring data of China Meteorological Administration.” China Weather Typhoon Network. https://typhoon.weather.com.cn/. Deposited 4 October 2024.
47
ECMWF, Data from: “Integrated forecasting system operational forecast.” ECMWF. https://apps.ecmwf.int/archive-catalogue/?class=od&stream=oper. Deposited 4 October 2024.
48
ECMWF, Data from: “Integrated forecasting system single-column data assimilation.” ECMWF. https://apps.ecmwf.int/archive-catalogue/?class=od&stream=scda. Deposited 4 October 2024.
Information & Authors
Information
Published in
Classifications
Copyright
Copyright © 2025 the Author(s). Published by PNAS. This open access article is distributed under Creative Commons Attribution-NonCommercial-NoDerivatives License 4.0 (CC BY-NC-ND).
Data, Materials, and Software Availability
The GridSat-B1 dataset is from www.ncei.noaa.gov/products/gridded-geostationary-brightness-temperature (40). The ERA5 reanalysis data are available from https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=form (42) and https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-pressure-levels?tab=form (43). The CMA best track data are from https://tcdata.typhoon.org.cn/zjljsjj.html (45). The CMA operational data are available from https://typhoon.weather.com.cn/ (46). The IFS-Operational forecast data are available from https://apps.ecmwf.int/archive-catalogue/?class=od&stream=oper (47) and https://apps.ecmwf.int/archive-catalogue/?class=od&stream=scda (48).
Submission history
Received: August 1, 2024
Accepted: December 6, 2024
Published online: January 21, 2025
Published in issue: January 28, 2025
Keywords
Acknowledgments
This work was supported by the NSFC Innovative Group Grant (No. 42221005), the National Natural Science Foundation of China (grant nos. U2006211, 42376175, 42090044, and 42076200), the Strategic Priority Research Program of the Chinese Academy of Sciences (grant no. XDB42040401), the China-Portugal Xinghai “Belt and Road” Joint laboratory and joint research on new air and sea technologies (2022YFE0204600), and the Postdoctoral Fellowship Program of CPSF (GZC20241741).
Author contributions
C.W. and X.L. designed research; C.W. performed research; C.W. and N.Y. analyzed data; and C.W. wrote the paper.
Competing interests
The authors declare no competing interest.
Notes
This article is a PNAS Direct Submission.
Authors
Metrics & Citations
Metrics
Altmetrics
Citations
Cite this article
Advancing forecasting capabilities: A contrastive learning model for forecasting tropical cyclone rapid intensification, Proc. Natl. Acad. Sci. U.S.A.
122 (4) e2415501122,
https://doi.org/10.1073/pnas.2415501122
(2025).
Copied!
Copying failed.
Export the article citation data by selecting a format from the list below and clicking Export.
View Options
View options
PDF format
Download this article as a PDF file
DOWNLOAD PDFLogin options
Check if you have access through your login credentials or your institution to get full access on this article.
Personal login Institutional LoginRecommend to a librarian
Recommend PNAS to a LibrarianPurchase options
Purchase this article to access the full text.