Decadal Prediction of Location of Tropical Cyclone Maximum Intensity Over the Western North Pacific

Attaining skillful decadal predictions for the western North Pacific tropical cyclone (TC) emerges as a formidable challenge, mainly stemming from the limited prediction skills of Pacific Decadal Variability (PDV) within the state‐of‐the‐art models. Assessing sixth Coupled Model Intercomparison Project models' retrospective predictions finds that the predictability of PDV transcends the expectations set by raw forecasts featured by constrained temporal skills and low signal‐to‐noise ratio. Employing a refined approach, we selectively identify the models that capture the diverse phases of PDV and subsequently adjust their variances. This tailored approach yields a compelling concordance between the predicted PDV and observation in phases and variances. Anchored in the heightened prediction skill of PDV, we establish a sophisticated statistical model adept at predicting the latitude of TC's lifetime maximum intensity (LMI). The near‐term prediction indicates a sustained poleward migration of LMI latitude by 1.53° during 2020–2027, increasing subtropical East Asia's TC‐related disaster vulnerability in the coming decade.


Introduction
Tropical cyclones (TCs) stand among the most catastrophic natural events, contributing to a third of fatalities and economic damages resulting from weather-, climate-, and water-related calamities (World Meteorological Organization, 2019).This proportion is notably amplified in East Asia due to its proximity to the world's most active TC basin, the western North Pacific (WNP) (Mendelsohn et al., 2012;Peduzzi et al., 2012;Q. Zhang et al., 2009).Unquestionably, skillful prediction of TC activity over the WNP would greatly benefit people in the Pacific coastal regions.Predictions ranging from synoptic to interannual timescale have provided increasingly useful information to support disaster planning and migration (Klotzbach et al., 2019).
On a longer time scale, near-term predictions, also known as decadal predictions with time scale ranging from years to decades, bridging the gap between seasonal forecasts and long-term projections, are of great socioeconomic value (Kushnir et al., 2019;Wang et al., 2018).However, currently, decadal prediction for TC activity has only been conducted for the North Atlantic basin due to the high predictability of Atlantic multidecadal variability (Caron et al., 2018;Smith et al., 2010), a main driver for decadal variability of TCs in the North Atlantic.The decadal prediction for the WNP TC activity has been scarcely investigated (Knutson et al., 2020).cyclone (TC) activity depends on Models' skills to predict Pacific Decadal Variability (PDV), which is limited • We substantially improved PDV's hindcast skill and the refined approach predicts a PDV negative phase in the near-term (2020)(2021)(2022)(2023)(2024)(2025)(2026)(2027) • The predicted PDV phase implies a poleward shift of location of TC's lifetime maximum intensity, increasing subtropical East Asia's TC-related risk

Supporting Information:
Supporting Information may be found in the online version of this article.
It has been found that Pacific Decadal variability (PDV) is the foremost driver for decadal variability of the WNP TC activity (Liu & Chan, 2013;Sobel et al., 2023).Practically, the location of TCs reaching their lifetime maximum intensity (LMI) over the WNP has experienced a robust poleward migration since 1980 (Kossin et al., 2014) due to the phase shift of PDV in the late 1990s (Song & Klotzbach, 2018;Zhao et al., 2022), which has caused increasingly hazard exposure and mortality risk for subtropical East Asia (Lee et al., 2021;Park et al., 2011;Wu et al., 2022).Therefore, how PDV evolves in the near future would greatly constrain the WNP TC activity.
Decadal prediction is complex since it is influenced not only by boundary conditions, but also by initial conditions (Boer et al., 2016;Meehl et al., 2021).Current initialization predictions show high skill in retrospective predicting sea surface temperature (SST) decadal variability in the North Atlantic, Indian Ocean and Southern Hemisphere (Simpson et al., 2019;Smith et al., 2020), but low skill in the Pacific.For instance, the initialized state-of-the-art coupled models show skillful predicting of PDV only with a lead time of one to 2 years, and the prediction skill decays rapidly as the lead time increases (Choi & Son, 2022;Kim et al., 2012;Wiegand et al., 2019).Currently, decadal prediction of PDV has hit a bottleneck, greatly limiting the skill in decadal prediction of the Pacific climate (Boer & Sospedra-Alfonso, 2019;Kim et al., 2014).
This study aims to assess PDV predictability in the current initialized models and to improve the predictive skill through an innovative post-processing method.It is found that PDV's prediction skill surpasses raw models' counterparts, enabling the development of an adept statistical model for predicting the WNP TC LMI latitude.Near-term predictions reveal a continual poleward migration of WNP TC LMI latitude in the coming decade, heightening subtropical East Asia's susceptibility to TC-related disasters.

Data and Method
TC best track data are obtained from the Joint Typhoon Warming Center (JTWC).TCs are defined as its maximum sustained wind speed greater than or equal to 35 knots.Location of LMI is defined as the latitude where the TC first reached its maximum intensity.Here we focus on the main WNP TC season from May to October, when about 80% of WNP TCs are formed.The results are independent of the season selection.Monthly SST data are from the National Oceanic and Atmospheric Administration (NOAA) Extended Reconstructed SST version 5 (ERSST v5) data set (Huang et al., 2017).PDV index is defined as the difference in SST anomalies between the eastern tropical Pacific (10°S-6°N, 110°W-160°W) and the North Pacific (30°N-45°N, 145°W-180°W) (Dong et al., 2014), and it is generally similar with PDO or IPO index (Henley et al., 2015;Mantua et al., 1997).The global warming (GW) index is defined as the mean SST between 40°S and 60°N (Zhao et al., 2022).
The decadal prediction experiments are obtained from the Decadal Climate Prediction Project (DCPP) of sixth Coupled Model Intercomparison Project (CMIP6) (Boer et al., 2016).It is a set of retrospective decadal forecasts (known as hindcasts) containing 166 ensemble members initialized in October, November, or January of every year over 1960-2018 (Table S1 in Supporting Information S1).Each member integrates for 10 years.Here the decadal prediction is expressed as an 8-year average for 2nd-9th years of prediction, which can exclude the influence of season-interannual forecast skills on decadal forecasting (Smith et al., 2020).The 8-year rolling mean in the predictions is validated with the 8-year running mean in the observations.Since the simulated mean state may differ among models, the mean states for individual models and observation have been first removed to eliminate their difference.Accordingly, predictions are expressed as anomalies relative to their own climatology (Boer et al., 2016).The multi-model ensemble mean is calculated by the unweighted mean of all ensemble members.All model data and observational data are interpolated to a uniform 2.5°× 2.5°grid before comparison.
Three indices were assessed to measure predict quality (Text S1 in Supporting Information S1), the anomaly correlation coefficient (ACC), mean-squared skill score (MSSS) and the ratio of predictable components (RPC) (Smith et al., 2020).To understand the sources of decadal predictions skills, we use the uninitialized CMIP6 historical simulations of the corresponding models over 1962-2014 (132 ensemble members; Table S1 in Supporting Information S1) to estimate the contribution of uninitialized components (external forcing) to the predicting skill (Text S2 in Supporting Information S1) in the decadal hindcast experiment (Sospedra-Alfonso & Boer, 2020).
A variance adjustment procedure is adopted to align the variance of the predicted signal in the raw PDV time series with the observed variance (Text S4 in Supporting Information S1) (Hu & Zhou, 2021;Smith et al., 2020).

LMI latitude prediction is based on the regression relationship equation (Text S5 in Supporting Information S1).
By considering the uncertainty of regression coefficients, model predictions, and high-frequency noise, the original prediction system consisting of 38 members of the four selected models was expanded into a prediction set of 38,000 members (Text S6 in Supporting Information S1).Two of the four selected models (MIROC6 and EC-Earth3) contain 20 members to provide SST predictions starting in 2018, so the LMI latitude predicted from 2018 contains 20,000 members.The significance test is based on the nonparametric block bootstrap approach (Text S3 in Supporting Information S1) (Smith et al., 2020).

Evaluating Model's Hindcast Skill
GW and PDV are two dominant factors modulating the WNP TC LMI latitude (Zhao et al., 2022), thus, skillful prediction of GW and PDV is a prerequisite for decadal prediction of the WNP TC LMI latitude.Here we first evaluate the predicting skill of GW and PDV in the DCPP experiments.As expected, the current state-of-the-art coupled models can well predict the persistent warming trend since 1960, with the ACC exceeding 0.96 in all 11 models from 1962 to 2020 (Figure S1 in Supporting Information S1).However, these models show distinct skill in predicting PDV index with ACC ranging from 0.05 in the CESM1-1-CAM5-CMIP5 model to 0.57 in the IPSL-CM6A-LR model (Figure 1).Why do models exhibit different levels of skill in predicting PDV?The spatial distribution of ACC skill for the top-four and bottom-four models are examined (Figures 2a and 2b).Significantly positive skill can be found in the most of north Pacific except the central Pacific in both groups.Nonetheless, ACC skills in the key regions defining PDV of the top-four models are higher than those in the bottom-four (Figure 2a).The ACC skill can be further decomposed into the contribution from the initialized and uninitialized components (Sospedra-Alfonso & Boer, 2020).It can be found that better skill of the top-four models in subtropical Pacific mainly comes from their more reasonable response to external forcing, while their better skill in the equatorial central Pacific is attributed to the model initialization (Figures 2c and 2d).
The multi-model ensemble mean (MME) is common way to reduce biases from different models and members and to obtain better prediction (Reichler & Kim, 2008).While the MME predicting skill of all 166 members for the GW index is good due to their high consistency in 11 models (Figure S1 in Supporting Information S1), the MME hindcasted PDV shows an insignificant ACC skill of 0.34, which is lower than that in six of 11 models.The low skill of PDV indicates the MME cannot eliminate the biases across models, and it is essential to exclude the models with low skill before constructing the MME.Here, top-four models including BCC-CSM2-MR, EC-Erath3, IPSL-CM6A-LR, and MIROC6 models with 38 members, whose ACC are significant at 90% level,

Variance Adjustment for PDV
Besides the temporal phase, the magnitude of the predicted anomaly is important as well.The MSSS for the predicted GW index reaches 0.97 (Figure S2 in Supporting Information S1), suggesting the reasonable skill of the model in predicting the magnitude of GW index.However, a low MSSS (0.19) with a high ACC (0.65) is obtained for the predicted PDV, indicating a possible underestimation of the predicted amplitude in the MME (Figure 3).Variables in climate models can be decomposed into the potentially predictable component (referred to as signal) and the inevitably unpredictable noise (referred to as noise) due to the chaotic nature of atmosphere.The underestimation indicates that each ensemble member contains too much noise and thus a low signal-to-noise ratio in the model.Particularly, the total variability of PDV measured by standard deviation ranges from 0.24 to 0.44 for individual members, which is generally consistent with the observation (0.37).Nevertheless, the standard deviation of the MME decreases sharply to 0.15, which means that the variance of the observed predictable signal is approximately 2.6 times greater than that of the model.The errors of the signal-to-noise ratio can be quantified by the RPC in models (Smith et al., 2020).For a perfect prediction system, the predictable component in the model and observation should be equal, and thus RPC should be equal to 1. RPC greater than 1 denotes that ratio of predictable component is lower in models than in observations.The calculated RPC for the MME is 1.40, indicating a low signal-to-noise ratio in model, and thus skill of the models to predict PDV is underestimated.
To correct the low signal-to-noise ratio in the original prediction of PDV, a variance adjustment technique (Hu & Zhou, 2021;Smith et al., 2020) is conducted on the predictable signal and unpredictable noise in the prediction (Figure 3).Indeed, the selected model prediction (Figure 3a, ACC = 0.65, MSSS = 0.19) outperforms the persistence prediction (ACC = 0.29, MSSS = 0.51) and the historical simulation prediction (ACC = 0.22, MSSS = 0.18).After variance adjustment, the spread in individual member is greatly reduced and the amplitude  becomes larger and much closer to the observation (Figure 3b).Accordingly, the MSSS for the adjusted prediction increases to 0.38 from 0.19, indicating validity of the MME in the eliminating model errors.Moreover, ACC of the adjusted prediction increases to 0.75 (p < 0.01), and the evolution is very similar to the observation (Figure 3b).Particularly, the adjusted prediction can reasonably reproduce the positive phase of the early 1970s to the late 1990s and the negative phases in the rest period, which is much improved compared to the pre-adjusted prediction (Figures 3a and 3b).The adjustment not only improves the ensemble mean skill but also the skill of the individual member.For the pre-adjusted prediction, the proportion of members with ACC values greater than 0.5 is only 2.63% and none of members with positive MSSS.After the adjustment, the ACCs of all members exceed 0.5, and 18.4% of members with MSSS greater than 0 (Figures 3e and 3f), indicating a notable improvement of the adjustment.

Decadal Hindcast and Quasi-Real-Time Prediction of the WNP TC LMI Latitude
The LMI latitude of the WNP TCs exhibits notable decadal variability, which is found to be linked to GW and PDV (Zhao et al., 2022).Since PDV is closely linked to ENSO (Newman et al., 2016;Y. Zhang et al., 1997), the influence of decadal component of ENSO on TC LMI latitude has been largely included.Accordingly, the regressed LMI latitude using the PDV and GW indexes as variables agrees well with the observation with a correlation coefficient reaching 0.85 (Figure 4, p < 0.01), suggesting about 70% of the decadal variance of the WNP TC LMI latitude can be interpreted by GW and PDV together.This tight linkage lends further confidence to predict the WNP TC LMI latitude due to the potential predictability associated with the high predicting skills of GW and PDV (Figures 2 and 3).Taking the original GW index and the adjusted hindcast PDV index in the DCPP experiment into the established regression equation instead of the observed indexes, the hindcast can reasonably reproduce the observed evolution of LMI latitude with a significant ACC of 0.86 (Figure 4, p < 0.01), indicating the high-level skill of this predicting strategy.
How the WNP TC LMI latitude will change in the coming decade means a lot for decision-makers in the context of climate risk management (Kushnir et al., 2019).Based on the available 20 members initialized from 2018, 20,000 members are generated through bootstrapping sample (Text S6 in Supporting Information S1) for the predictions of TC LMI latitude.Since most available models start from 2018 and integrate for 10 years, generating the available predictions of 2019-2028 and we focus on prediction for the next 2-9 years, and thus yielding the period of 2020-2027 for the near future.The prediction suggests that PDV will shift to the negative phase during 2020-2027 (Figure 3).It can be seen that the variance-adjusted prediction of PDV reduces the uncertainty by about 20% compared to the pre-adjusted prediction (Figures 3c and 3d).The prediction indicates a negative PDV anomaly of 0.79 K during 2020-2027, which is strongest negative phase of PDV compared to the historical records (Figure 3).The negative phase of PDV can lead to a northward shift in TC genesis and thus a poleward migration of TC LMI latitude.Moreover, models predict a persistent GW (Figure S2 in Supporting Information S1), favoring a northward shift the WNP TC track (Zhao et al., 2022).Therefore, the possible strongest negative phase of PDV, in addition to the persistent GW, leads to the continual poleward moving of WNP TC LMI latitude by 1.53°in the coming decade (Figure 4).The possibility of exceeding the historical extreme high TC LMI latitude (2012-2019) exceeds 85% (Figure 4), suggesting a very possible unprecedent TC activity in subtropical East Asia and a giant challenge for the disaster migration therein.

Summary and Discussion
Decadal prediction stands as a pivotal frontier in addressing climate change concerns.Notably, the challenge of predicting WNP TC activity on a decadal scale has been complicated by the limited predictability of PDV, a primary driver of WNP TC variations.However, the conventional approach of relying on large ensemble means proves suboptimal due to systemic biases in certain models.To surmount this, we scrutinize retrospective climate model predictions spanning six decades, revealing that PDV's predictability far exceeds the modest temporal accuracy and signal-to-noise ratio evident in raw model outputs.Our approach involves the selection of PDVphase-capturing models, followed by the calibration of modeled variances to closely align with observations.Harnessing the heightened PDV predictability, we devise a statistical model adept at predicting TC latitude at LMI, displaying robust skill with a correlation coefficient of 0.86 from 1962 to 2020.Near-term predictions portend an unprecedented poleward shift in WNP TC activity due to the predicted negative PDV phase and persistent GW, posing significant challenges for disaster management in subtropical East Asia during the current decade.
Validation of the 2020-2027 prediction finds partial substantiation through available data from 2020 to 2022.Notably, current PDV and LMI latitude anomalies are generally consistent with the observation, characterized by a negative PDV phase accentuated by consecutive La Niña events in 2020-2022 and a poleward LMI latitude migration (Figures 3 and 4).Recent occurrences such as typhoon Doksuri (2305) and typhoon Khanun (2306) causing extensive devastation in subtropical East Asia further align with our predictions.To the best of our knowledge, it is the first attempt to construct the decadal prediction for TC activity, which establishes a promising avenue for skillful decadal PDV prediction and bridges a critical gap in WNP TC activity prediction.Furthermore, given the pivotal role of PDV in decadal climate prediction within the Pacific region, this study holds crucial implications for forecasting the broader Pacific climate on a decadal scale.These findings have the potential to substantially enhance disaster prevention efforts across the pan-Pacific region, bolstering preparedness measures and overall resilience.
Although here we predict a robust poleward migration of TC LMI position due to the negative PDV phase and continual GW in the coming decade, one should note that the prediction uncertainty still exists due to this indirect strategy.It is more straightforward to examine changes in directly simulated TCs in the DCPP experiment, and collaboration with the high-resolution decadal prediction outputs that can resolve TCs deserves further study.Additionally, one may concern the possible sources of models' diverse skill in predicting PDV.Due to the similar external forcing conditions used in current climate models, the initialization scheme (Mochizuki et al., 2010) and model performance are the main sources of differences in model prediction skills.Some studies suggest that the dynamic adjustment of heat content anomalies under initial conditions in tropical regions is the key to the successful transformation of PDV phase in model prediction (Ding et al., 2013).In addition, due to insufficient understanding of decadal climate variability, the model may not include some important physical processes that are important for decadal variation in climate (Kushnir et al., 2019;Latif Mojib & Keenlyside, 2011).

Figure 1 .
Figure 1.Anomaly correlation coefficient (ACC) skill for predicting Pacific Decadal Variability index in climate models.The blue bar presents multi-model ensemble mean prediction and the red bars present the selected models whose ACC skills are significant at 90% confidence level.Numbers denote the ACC for individual models.

Figure 2 .
Figure 2. Decadal prediction skill for 2-9-year sea surface temperature.(a) anomaly correlation coefficient (ACC) skill for the ensemble mean prediction of (a) top-four models (38 members) (b) bottom four models (58 members) for 1962-2020, (c) the differences in ACC between the top-four models and bottom four model ((a, b)), and the difference contributed by the (d) initialized components (shadings) and uninitialized components (contours, ranges from 0.1 to 0.3 with an interval 0.2, bold purple contours denote 0.1), hatched and dot regions denote the values significant at the 90% confidence level.The blue boxes denote the regions defining the Pacific Decadal Variability index.

Figure 3 .
Figure 3. Decadal prediction for Pacific Decadal Variability (PDV) index.(a) The 2-9-year rolling mean PDV index from decadal hindcasts for 1962-2020 (blue line), the quasi-real-time predictions (red line) in the top-four models, and the corresponding observations (black triangles), the boxplot denotes the prediction for 2020-2027.The limits of the whisker represent the minimum and maximum, the limits of the boxes represent the 25th and 75th percentiles, and the horizontal line and circle inside the box are the median and the mean.The yellow triangles denote the observation for 2020-2022.Panel (b) same as in panel (a), but for the variance adjusted model hindcast and prediction.Anomaly correlation coefficient (ACC), mean-squared skill score (MSSS), and ratio of predictable components are shown in each panel.Shading in (a) and (b) show the uncertainty range.Probability distributions of predicted PDV in 2020-2027 for the (c) raw and (d) variance-adjusted forecast ensembles.The numbers in (c) and (d) show the median and standard deviation values (number in the brackets).(e) Hindcast ACC for the raw (left) and variance-adjusted (right) forecast ensembles, black values in abscissa denote the probability in % that the ACC is greater than 0.5.(f) Hindcast MSSS range for the raw (left) and varianceadjusted (right) forecast ensembles, black values in abscissa denote the probability in % that the MSSS is greater than 0. The gray dash lines and numbers in (e) and (f) show the ACC and MSSS of persistence hindcast.

Figure 4 .
Figure 4. Decadal hindcast and quasi-real-time prediction for the western North Pacific tropical cyclone lifetime maximum intensity (LMI) latitude.The 2-9-year rolling mean LMI latitude from decadal hindcasts for 1962-2020 (blue line), the quasi-real-time predictions (red line) in the top-four models, the regression predictions (green line) and the corresponding observations (black triangles).The shading shows the uncertainty range.The yellow triangle denotes the observation for 2020-2022.The boxplot denotes the prediction for 2020-2027.The limits of the whisker represent the 5th and 95th percentiles, the limits of the boxes represent the 25th and 75th percentiles, and the horizontal line and circle inside the box are the median and the mean.The right panel shows possibility distribution of the predicted LMI latitude in 2020-2027.The gray dash line denotes the historical extreme LMI latitude of 2012-2019.The percentage denote the possibility of the prediction exceeding the extreme LMI latitude of 2012-2019.Anomaly correlation coefficient are shown in the panel.