A New Method for Predicting Hurricane Rapid Intensi�cation Based on Co-occurring Environmental Parameters

8 Tropical cyclones (TCs) that undergo Rapid Intensiﬁcation (RI) can pose 9 serious socioeconomic threats and can potentially result in major dam-10 aging impacts along coastal areas. Considering the complexity of various 11 physical mechanisms that play a role in RI and its relatively low probabil-12 ity of occurrence, predicting RI remains a major operational challenge. In 13 this study, we propose a simple deterministic binary classiﬁcation model 14 based on the co-occurrence of environmental parameters (MCE) to pre-15 dict an RI event. More speciﬁcally, the model determines the possibility 16 of RI based on a simple count of the number of environmental predic-17 tors deemed favorable and unfavorable. We compare our model results to 18 logistic regression (LR) and decision tree (DT) models, well-trained using 19 the same set of environmental predictors. Results reveal that at an RI 20 threshold of 30kt, the MCE exhibits a critical success index (CSI) score of 21 0.233 which is 14% higher than DT and LR model performances. When 22 tested at multiple RI thresholds, the MCE displays relatively higher skill 23 scores across multiple metrics. By simultaneously evaluating the favora-24 bility of predictors, the MCE is able to comparatively reduce the number 25 of false alarms predicted when certain predictors are unfavorable towards 26 RI. Interpreting these model results to gain a physical understanding 27


Introduction
The development and landfall of hurricanes or tropical cyclones (TCs) can potentially result in significant damages for coastal regions.The mechanisms behind TC intensification and weakening are complex, making TC intensification difficult to predict accurately.Even more so, TC rapid intensification (RI), defined as an instance when the storm's maximum sustained surface wind speed increases by 30 kt or more over a 24-hour period (Kaplan and DeMaria, 2003), is substantially harder to predict given the low probability of occurrence.Historically, almost all TCs that have attained Category 4 or 5 strength underwent RI during their lifetimes (Kaplan and DeMaria, 2003), further emphasizing the need to improve RI forecasts.For instance recently, Hurricane Ian underwent RI and made landfall over Florida peninsula as a strong and catastrophic Category 4 hurricane in September 2022 (Espinel et al, 2022).Similarly, in August 2021, Hurricane Ida underwent RI and struck New Orleans, Louisiana as a Category 4 storm and inflicted substantial damages (Zhu et al, 2022).In the operational RI forecasting space, a suite of dynamical models and statistical methods have been employed for RI prediction.With modest advancements in RI prediction over time, there lies considerable room for improvement (DeMaria et al, 2021) as predicting RI events remains a high priority for the National Hurricane Center (Rappaport et al, 2009), (Kaplan et al, 2010).
A multitude of challenges must be dealt with when predicting RI.First, there is uncertainty surrounding the underlying physical processes of a TC undergoing RI.Numerous studies (Kaplan et al, 2010), (Kaplan et al, 2015) have explored which select few environmental predictors favor RI formation.
However, the relative significance of these environmental predictors for RI remains unclear.More confounding is that no one particular set of environmental conditions can guarantee an RI event (Rozoff et al, 2015).In addition, the availability of RI data is limited compared to non-RI data.Consequently, training of RI prediction models can be severely hampered by datasets.Common issues that RI prediction models face are poor probabilities of detection (POD) and high false alarm ratios (FAR) (DeMaria et al, 2021).Despite these challenges, numerous studies have addressed this problem to gain a better physical understanding of the phenomenon, and consequently, improve our ability to predict the possibility of RI.
One of the first notable statistical models developed for RI forecast was the Statistical Hurricane Intensity Prediction Scheme (SHIPS) which used multiple linear regression techniques to predict intensity changes for Atlantic TCs (DeMaria and Kaplan, 1994Kaplan, , 1999;;DeMaria et al, 2005).Kaplan and DeMaria (2003) extended this to implement an RI-specific, threshold-based probabilistic prediction scheme called SHIPS -Rapid Intensification Index (SHIPS-RII).This was later expanded upon by Kaplan et al (2010) to include the North Pacific basin and different RI thresholds (25, 30 and 35 kt).To improve RI prediction skill, Kaplan et al (2010) designed the SHIPS-Linear Discriminant Analysis model (SHIPS-LDA) that performed an LDA (Daniel and Wilks, 2006)  Machine learning (ML) techniques are increasingly being used for TC intensity prediction, including RI, with the aim of capturing the non-linear relationships between environmental predictors and storm behavior.Initial ML efforts started with Rozoff and Kossin (2011) that introduced a logistic regression (SHIPS-LR) and a naive Bayesian model (SHIPS-NB).These models are used with the SHIPS-LDA to create a consensus SHIPS model (SHIPS-C) (Kaplan et al, 2015) that is used by the NHC in an operational setting.SHIPS-LR employs the logistic regression technique which is commonly used for binary predictands.It assigns a regression weight to each individual predictor which is summed up with predictor values to output a probabilistic prediction of RI.The Naive Bayesian equation (Kossin and Sitkowski, 2009) in SHIPS-NB uses Bayes theorem to predict the conditional probability of an RI event.There have been many studies that investigate the intensification of TCs using association rule algorithms (Yang et al, 2007(Yang et al, , 2008(Yang et al, , 2011)), data mining techniques (Yang, 2016) classification and regression trees (Wei and Yang, 2021), decision trees (Zhang et al, 2013), (Kim et al, 2021), long shortterm memory models (Yang et al, 2020), multi-layer perceptron models (Xu et al, 2021) and ML ensembles (Mercer and Grimes, 2017;Su et al, 2020) that show skill in predicting RI.
Across the multitude of RI prediction approaches, there is little work that explores particularly how environmental parameters can concurrently contribute to a potential RI event and how that can be leveraged into an RI prediction model.Most models evaluate RI probability using methods which tend to be influenced by a limited set of strong predictors rather than looking at whether the large-scale environment is comprehensively conducive across the board.Herman and Schumacher (2018) explored the advantages and disadvantages of logistic regression (LR) models for forecasting extreme precipitation, which has a low probability of occurrence similar to RI.They suggest that a shortcoming that LR faces is that predictor regression coefficients are applied uniformly across data samples.For example in models like SHIPS-LR and SHIPS-LDA, if certain predictors independently can not largely influence RI, but in conjunction with other predictors can lead to high chance of RI, a trained model would still assign weaker coefficients to these predictors and stronger weights to the other predictors (Herman and Schumacher, 2018).This could lead to cases where a select few highly weighted predictors strongly influence the model's RI prediction, causing the model to overlook other predictors that might not favor RI, leading to a potential false alarm.
In this study, we propose a simple deterministic Model based on the co-occurrence of environmental parameters (MCE) to predict an RI event.
We compare our model results to logistic regression and decision tree based approaches and interpret these models' results to explore the potential dynamics of how simultaneously co-occurring environmental parameters can affect a possible RI event.A logistic regression model was chosen for comparison, considering its use in the SHIPS-Consensus model and its contrasting method that doesn't explicitly evaluate predictors concurrently to predict RI.The decision tree model was also used for comparison to explore how environmental predictors are evaluated in the hierarchical structure of the model's decision rules.(Zhang et al, 2013).The main objectives of the study are as follows: 1. To demonstrate the utility of the MCE for predicting RI events using cooccurring environmental parameters.Most RI prediction studies (for e.g (Kaplan et al, 2010), (Mercer and Grimes, 2017), (Yang, 2016)) were conducted at the basin scale; however, this study aims to look at how considering the co-occurrence of environmental parameters may improve RI prediction at the global scale.Further, this can lead to a more general understanding of the underlying physical mechanisms of TC RI.In addition, inclusion of all the SHIPS basins increase the amount of available training and testing data for a more robust and informed model.

Environmental Predictor Selection
In this study, we focus our efforts on 5 environmental predictors from the SHIPS global dataset that were used in Kaplan et al (2015) for the revised SHIPS-RII: Potential Intensity (POT), Vertical Wind Shear (SHRD), Relative Humidity at 700 hPa (RHLO), Divergence at 200 hPa (D200) and Ocean heat Content (OHC).Kaplan et al (2015) showed that these variables exhibited statistically significant differences at the 99.9% level using a twosided Behrens-Fisher t-test (Dowdy and Wearden, 1991) between the RI and non-RI data samples for RI thresholds of 25, 30 and 35 kt.Higher values of POT, RHLO, D200 and OHC positively correlate with higher chance of RI whereas lower values of SHRD favor RI (Kaplan et al, 2015).We limit our predictor selection for simplicity and ease of understanding.However, the technique presented here could be extended to include different and larger selections of environmental predictors.The average RI and non-RI values of the environmental predictor data used in this study is shown in Figure 1.
Predictors are taken either at t=0 of the storm or a temporal average from t=0 to t=24hr.A table outlining the specific SHIPS predictors used to derive the environmental predictors is shown here 1. POT is the potential intensity calculated following the method described in Kaplan and DeMaria (2003) by subtracting the intensity of the current storm (VMAX), from the average of the maximum potential intensity (VMPI) from t=0 to t=24.OHC is derived from the NCODA analysis (denoted as NOHC in the SHIPS database).Since at the time of the study, NOHC was not available for the Western North Pacific, Northern Indian and Southern Ocean basins, the RHCN derived from satellite altimetry data is used in place of NOHC for these basins.If RHCN data is missing, PHCN, which is the estimated ocean heat content from climatology and the current SST anomaly, is designed to fill in for RHCN as per the SHIPS documentation.
Any non-existent variables for the predictors are removed from the dataset.
All overland TC locations are removed from the analysis.Further, those TC locations over water that had a landfall in the -12hr to 24hr time frame are also removed to ensure that our results are not contaminated by land effects.Furthermore, only those locations where a storm is at least of Tropical Storm (maximum sustained surface winds above 34 kt) strength are considered.

Data pre-processing
A scaling method similar to that in Kaplan et al (2010) is used to scale each predictor's values from 0 to 1.In this study, a scaled value of 0.0 is assigned to the dataset's minimum (maximum) value of POT, RHLO, D200, OHC (SHRD) and a scaled value of 1.0 is assigned to the dataset's maximum (minimum) value of POT, RHLO, D200, OHC (SHRD).Values in between are interpolated linearly.In the resulting dataset, the closer the predictor's values are to 1.0, the more favorable the value is for RI.
We approach RI prediction in these models as a binary classification problem.To create the predictand dataset, each data sample is categorized as an RI event (1) or non-RI event (0) when the DELV variable from the SHIPS dataset,

Parameters (MCE)
The MCE method is intentionally kept simple to focus on the simultaneously co-occurring environmental parameters and gain insight into how large-scale environment's conduciveness can affect RI.To develop our model, the training dataset is used exclusively and we reserve the testing dataset to test model performance.For each environmental predictor, we use a threshold similar to Kaplan and DeMaria (2003) to determine whether the value of the predictor is favorable and a separate threshold to determine whether the predictor is unfavorable towards RI.Details on this classification method are given in Table 2 and the threshold values for each environmental predictor are outlined in Table 3.More details on the methods used to find the optimal thresholds are given in the Supplementary Information.
For each predictor that is favorable, the net favorable predictor count increases by 1.If a predictor is deemed as unfavorable, the net favorable predictor count is reduced by 1.If a predictor is neither favorable nor unfavorable, the count is not affected.This allows the MCE to account for potentially non-conducive predictors and use only the net favorable environmental predictors when predicting RI.Finally, if the net favorable predictor count exceeds a certain count threshold, the model predicts an RI event.

Decision Tree Classifier (DT)
A decision tree (DT) binary classifier was developed to compare RI prediction performance with the MCE models.The tree-like structure of the decision-making process evaluates the conduciveness of the environment by hierarchically checking whether environmental predictors meet certain thresholds set by a trained decision tree model.However, one difference between the MCE and DT model is that in training, the DT model can utilize certain environmental predictors more frequently in the decision rules over other predictors.In this instance, the model may be more influenced by certain predictors when predicting RI.The trained DT model outputs an RI probability value depending on whether this exceeds a pre-determined probability threshold.Similar to the LR, the GridSearchCV with LOYO cross validation method evaluated on CSI is used to determine the best performing model.
Additional details regarding the parameter space and grid search results are given in the Supplementary Information.

Model Evaluation Metrics
The purpose of using multiple evaluation metrics to test the various models is to obtain a comprehensive overview of model performance and to pinpoint how each model approaches the RI prediction problem.In binary forecasts where models predict an event or nonevent for each data sample, evaluation metrics largely comprise of elements from a 2x2 contingency table that compare observations to model forecasts.The table captures the number of true positives (TP) where the model forecast RI and RI was observed, false positives (FP) where the model forecast RI and RI was not observed, false negatives (FN) where the model did not forecast RI and RI was observed and lastly, true negatives (TN) where the model did not forecast RI and RI was not observed.The common forecast evaluation metrics used in this study that are derived from these elements are described in table 4.

Model Performance Analysis
We analyze the MCE's performance in predicting RI using the optimized thresholds found in Table 3. In, Figure 2, we first look at a case where the MCE predicts RI using a single set of thresholds, from column 1 in Table 3, to determine if the environmental predictor is favorable towards RI.In this case, we set a count threshold of 3 indicating that if a data sample has at least 3 simultaneously favorable environmental predictors, the MCE predicts RI.We use the resulting best MCE model derived from the methods outlined in the Supplementary Information.The results show that the MCE exhibits a CSI score of 0.17 with a POD of around 0.56 and a high FAR around 0.81.
In comparison, for the same model, if we set a count threshold of 4 favorable predictors in order to predict RI, the MCE exhibits an increase in CSI score to around 0.21.By ensuring a larger portion of the overall environment is favorable, the MCE demonstrates higher CSI attributed to a marked improvement in FAR outweighing the milder decline in POD.However, when we compare to the LR and DT models, the MCE's CSI scores are only comparable.The models' FAR scores are still large which is a common issue in RI prediction models.
We introduce a second set of thresholds into the MCE, specifically meant to determine if environmental predictors are unfavorable towards RI which can be found in column 2 in Table 3.By doing so, we see a further improvement of around 10% in the MCE's CSI score over LR and DT models.Given non-RI events are abundant in our dataset, the CSI metric ignores the models' easily correct non-RI predictions which can tend to over-inflate other metrics of skill that account for these TNs (Daniel and Wilks, 2006), thus giving a truer picture of model skill in predicting RI.
The number of unfavorable predictors negatively impact the net favorable predictor count, as described in the MCE methods section.In this instance, keeping the count threshold at 4 net favorable predictors ensures that there are only 2 scenarios in which the MCE predicts an RI event.This is when either 4 or 5 environmental predictors are favorable towards RI.However, an RI event is not predicted in cases where 4 environmental predictors are favorable and 1 environmental predictor is unfavorable, since the unfavorable predictor negatively impacts the net count of favorable predictors to reach the set threshold of 4. By ensuring the MCE does not predict RI when one or more unfavorable predictors are present, there is a 5% reduction in FAR.The MCE performs best when both sets of favorable and unfavorable thresholds are included to evaluate the data samples and ensuring 4 net favorable predictors are necessary to predict RI.The MCE exhibits a CSI score of 0.233, 14% higher than LR and DT models as well as a higher POD score and lower FAR score.

A Comparison of Model 2x2 Contingency Scores
A summary of the models' 2x2 contingency table scores are shown in Figure 3. Figure 3a represents the contingency scores for all of the samples in the testing dataset.Figure 3b represents the subset of test cases where the environment was unfavorable, meaning those data samples had less than 4 net favorable environmental predictors as determined by the MCE thresholds.In these samples, 248 of them were RI cases and 5027 of them were non-RI cases.
In the overall testing dataset in Figure 3a, around 9% of the cases were RI events, though when looking at only the testing cases in the unfavorable environment in Figure 3b, RI cases made up only around 5% of the dataset.
Of the observed RI events in Figure 3a, we see the MCE predict the highest TP RI events and the lowest number of FNs compared to the LR and DT models which can explain MCE's higher CSI score since TPs and FNs are more relevant for CSI.
In Figure 3b, where the environment is unfavorable, the MCE does not predict RI events at all, having no TPs nor any FPs.RI events do occur in the unfavorable environment though at significantly lower percentages.In an unfavorable environment, where observed non-RI events make up around 95% of the test cases compared to 90% in the overall dataset, the MCE accurately predicts significantly more non-RI events (TNs) compared to LR and DT models.
In addition, despite the environment being unfavorable, DT and LR models predicted 157 and 254 RI cases that were FPs.Whereas, the MCE has 0 FPs.This can indicate the importance of accounting for co-occurring environmental parameters when models predict RI, especially when the large scale environment can be unfavorable.More so, this outweighs the loss in the MCE's skill in predicting TPs, in comparison to the LR and DT model, which only predicted 50 and 23 TP cases respectively.

Model Sensitivity to Multiple RI Thresholds
We test the sensitivity of the MCE through added testing at multiple RI severity thresholds in addition to the 30 kt RI threshold.From the reserved testing dataset, the models were tested on samples that showed RI at thresholds of 25kt, 35kt, and 40kt.There are 718 (338, 230) RI cases and 5401 (5781, 5889) non-RI cases in the additional testing datasets based on a 25 (35,40) kt RI threshold.
Model performance results across RI thresholds are shown in Figure 4.
We see a common trend in model performance as we increase the RI thresholds.As we narrow down the testing data to see how the models perform for more severe cases of RI, the POD increases, FAR increases and overall CSI decreases.This could be attributed to the testing RI sample size decreasing at higher RI thresholds.In addition, cases of more severe RI tend to have even lower probabilities of occurrence and usually require highly favorable environmental predictors.It is likely that most models predict RI when predictors are highly favorable in these severe RI circumstances.This can explain the increase in POD and FAR across the three models.Interestingly though, across the increasingly severe RI cases, the MCE consistently shows higher CSI scores driven by a lower false alarm ratio compared to LR and DT models at the same RI thresholds.This indicates that for higher RI cases, in comparison to models like the LR and DT that are not explicitly dependent on co-occurring parameters, the MCE shows comparatively improved skill in RI prediction.
Results from further analysis into model performance using the aforementioned evaluation metrics are shown in Figure 5.The MCE consistently shows higher PSS, F1 and GSS scores and lower FAR scores compared to LR and DT models.This indicates that in spite of the MCE's simple threshold based decision making, the model consistently outperforms across multiple metrics of skill.
In Figure 5a, the MCE exhibits higher POD scores than the DT model across RI thresholds but in comparison to LR, the MCE has POD scores only on par.Interestingly, despite similar POD scores between MCE and LR models, the MCE exhibited higher PSS scores seen in Figure 5c.Since PSS can also be defined as the difference between the POD and probability of false detection (POFD) (Daniel and Wilks, 2006), this indicates the MCE had lower rates of POFD compared to LR.In other words, in comparison to LR across RI thresholds, the MCE had a lower rate of observed non-RI events being incorrectly forecast as RI events.PSS differs from the FAR metric which evaluates the fraction of predicted RI events that were actually observed non-RI events.Comparatively, the principle of the MCE is to consider each predictor simultaneously at equal proportions to evaluate the environment's favorability towards RI.In the framework of LR and DT, if a few environmental parameters are unfavorable for RI, it could still lead to a RI prediction if the other parameters are highly favorable.This could explain the increase in FPs for the LR and DT prediction models in cases where the environment is unfavorable, shown in Figure 3b.

Conclusions and Discussion
The aim of the study was to explore how accounting for co-occurring environmental parameters can improve RI prediction.We create a simple binary RI prediction model, one that solely depends on ensuring there are multiple favorable environmental predictors while accounting for those predictors that can be unfavorable.The MCE model predicts RI if there are at least 4 net favorable predictors out of the 5 listed environmental predictors used in this study.Overall, results show that the MCE outperforms a well-trained LR and DT model across multiple performance metrics.When evaluated at an RI threshold of 30kt, the MCE had a CSI score of 0.23 which is around 14% higher than LR and DT models.From the model 2x2 contingency scores, we can see that the MCE shows improved skill over the LR and DT models, primarily with more accurate RI predictions in the overall testing dataset and more accurate non-RI predictions in the unfavorable environment testing dataset.
When evaluated at higher RI thresholds, the MCE consistently exhibits lower FAR scores showing that concurring favorable environmental parameters can be particularly important for predicting cases of higher rates of intensification.
We can see from our feature analysis, the LR model assigns higher regression weights to certain predictors, such as wind shear, when predicting RI.
The DT model accounts for environmental predictors to hierarchically meet certain thresholds in the model rules.However, the model favors incorporating certain predictors like potential intensity and wind shear at a much higher percentage in the model's decision rules over other predictors.On the other hand, the MCE evaluates the predictors simultaneously without varying assigned predictor weights and considers the overall environment's favorability for RI.
Beyond improving prediction, these results can help improve our physical understanding of RI.They suggest that RI is more likely to happen when several environmental parameters align together rather than in situations where only one or two parameters are highly favorable.Further, the MCE suggests that the occurrence of both favorable and unfavorable environmental parameters plays an important role in distinguishing RI and non-RI events.
Given the limited number of predictors used in this study, future work can involve broadening the scope of environmental predictors investigated for use in the MCE.The model can also be expanded to include RI predictions at lead times longer than 24 hours and results can be analyzed at the individual basin level.Also, there is an opportunity to expand the MCE framework to account for the degree of favorability of an environmental predictor in the current threshold levels.Incorporating the MCE's simple method of evaluating large-scale environmental conduciveness to RI within other skillful RI models could potentially improve RI prediction performance.Additionally, the concept of the MCE can be applied to understand the physical underpinnings of other low-probability weather and climate extremes that tend to have substantial societal impacts.

Funding and Acknowledgements
to obtain individual weights for each environmental predictor based on their contribution to RI.The scaled predictor values with individual weights were summed up to obtain a probabilistic prediction of RI.More recently, DeMaria et al (2021) developed a model that has been in NHC's operational use since 2018.The Deterministic to Probabilistic Statistical model (DTOPS) converts deterministic intensity forecasts from SHIPS, statistical-dynamical, regional-dynamical and consensus models into probabilistic RI forecasts using a binary logistic regression.DTOPS demonstrates improved RI forecast skill.A more detailed report on the history and the comparative performance of the various NHC operational RI forecasting models discussed above is presented in DeMaria et al (2021).

2.
Interpreting the MCE results and the physical implications of co-occurring environmental parameters for TCs undergoing RI.In this study, environmental predictors from the Statistical Hurricane Intensity Prediction Scheme (SHIPS) database are used.The SHIPS database records the environmental conditions experienced by a TC in 6-hr timesteps from -12hr to 120hr relative to the current position.Data is obtained from model operational analyses and from satellite observations.For the environmental predictors chosen for this study, all fields are relative to the TC storm center determined by the NHC Best Track.The version of the SHIPS database that was used for the study at the time contained data from 1982 to 2020 for the Atlantic, Eastern Pacific and Central Pacific basins.The North Indian Ocean and Western North Pacific basins contained data from 1990 to 2020, while the Southern Hemisphere contained data from 1998 to 2020.
which represents the intensity change (kt) in maximum sustained surface wind speed from t=0 to t=24h, is above a set RI threshold of 30 kt.The resulting predictor and predictand datasets are split into training and testing datasets for model optimization and evaluation, respectively.The data samples from the last 5 years (2016-2020) from all basins are reserved so the models can be tested on unseen data.The remaining data samples are arranged chronologically and are used for model training purposes.In the training dataset, there were 1244 RI cases and 14877 non-RI cases.The testing dataset had 502 RI cases and 5617 non-RI cases.
was developed alongside the MCE to compare RI prediction performance.Logistic regression is used in the SHIPS-RII consensus model for NHC operational use.Since the logistic regression model does not explicitly depend on co-occurring environmental parameters, a comparison of the LR model with MCE will provide useful insight into the physical nature of RI.The logistic regression model is fit on the training dataset and produces a RI probability value.Depending onwhether the probabilistic output exceeds a pre-determined probability threshold, the LR model predicts an RI event.To determine the best performing model, the GridSearchCV method with Leave-One-Year-Out (LOYO) cross validation evaluated on critical success index (CSI) is used for tuning the model hyperparameters and finding the optimal probability threshold.More information on CSI is given in Table4.Details regarding the specific parameter search are included in the Supplementary Information.
An analysis of the environmental features for each model, and their relative significance, is discussed below.The feature importance scores determined by the DT model and the linearly scaled feature weight scores determined by the LR model are in Table 5.They show that certain predictors play a relatively more important role in RI prediction.For example, the DT model places high importance on POT and SHRD when evaluating a data sample, indicating POT and SHRD are considered more often in the decision rules set by the model.Because a larger percentage of the DT decisions are determined by the POT and SHRD values, they hold larger influence when the model predicts RI.Similarly, in the LR model, SHRD has a considerably larger weight compared to the other predictors.This indicates that with lower values of SHRD, the odds ratio for RI is increased by a larger magnitude than the other environmental predictors.Hence an increased SHRD predictor value would affect the resulting RI probability value returned by the LR model by a larger percentage.The LR model may attribute a higher RI probability to a non-RI data sample despite the data sample having unfavorable values of the other predictors.
A.N., K.B., W.X. and L.R.L. are supported by the Office of Science (BER) of the U.S. Department of Energy as part of the Regional and Global Model Analysis (RGMA) program area.A.N. and K.B.

Fig. 1
Fig. 1 Bar chart of average values of environmental predictors RI samples (orange) and non-RI samples (blue) with error bars depicting the standard deviation

Fig. 2
Fig.2Performance diagram summarizing multiple performance metrics of decision tree (star), logistic regression (triangle), and MCE (circle).Diagram compares the MCE-F using a count threshold of 3 favorable predictors (green), MCE-F using a count threshold of 4 favorable predictors (red) and MCE-U that utilized additional unfavorable thresholds (purple).Models were tested using an RI threshold of 30 kt. x axis shows Success Ratio = (1-FAR).POD on y axis.Contour lines show CSI scores.Dotted diagonal line represent bias scores

Table 1
also acknowledge support from NOAA's Climate Program Office, Climate Monitoring Program Predictors used in this study and the corresponding SHIPS predictors used to derive them.

Table 2
Tabledepictsthe workflow of the MCE to determine RI event.For each predictor, dependent on where its value falls relative to the specified thresholds, the predictor is classified as favorable, neutral or unfavorable.In each of these cases, they can add to, not affect or subtract from the net favorable predictor count.The net favorable predictor count ultimately decides how the model predicts RI dependent on whether the count exceeds a set number of co-occurring environmental parameters.Table 3 Environmental predictors' favorable and unfavorable threshold non-scaled values used by MCE.