Pacific subsurface ocean temperature as a long-range predictor of South China tropical cyclone landfall

Tropical cyclones (TCs) are among the most damaging natural hazards, and there is strong demand for skillful seasonal forecasts of landfall, especially for the highly exposed South China coast. Existing operational forecasts for the region are typically issued in March/April (near-season) and are often at basin scale rather than focused on landfall. Western North Pacific TC variability is linked to ENSO, though contemporaneous correlations with basin counts are weak; delayed impacts following strong El Niño events have been reported. Current South China landfall forecasts rely on some level of ENSO SST predictability and thus cannot be issued far in advance. Subsurface ocean temperature has long been recognized as important for seasonal prediction and is assimilated in coupled models, but has rarely been used to predict atmospheric phenomena at lead times longer than about 3 months. This study investigates whether equatorial Pacific subsurface temperature anomalies can enable year-ahead prediction of South China TC landfalls and proposes a subsurface index (subNiño4) to achieve this.

Prior studies document relationships between WNP TC activity and ENSO, including changes in large-scale circulation and genesis locations. Seasonal landfall prediction schemes for South China commonly exploit ENSO-related SST indices and are issued shortly before the TC season. Subsurface temperature has been utilized for seasonal prediction in other contexts, such as the Indian Summer Monsoon and Atlantic hurricane activity, and is central to ENSO theory and prediction (e.g., recharge–discharge and delayed oscillator frameworks). Observations show subsurface heat content can precede surface ENSO signals by up to ~15 months, helping to overcome the spring predictability barrier in ENSO forecasts. However, leveraging regional subsurface signals to predict atmospheric impacts at year-long lead times has been limited prior to this work.

• Data: South China TC landfalls primarily from CMA best track data within IBTrACS v04r00, with additional agency data from JTWC, HKO, and JMA (via IBTrACS). A South China landfall event was defined as the first timestep a TC center is within 200 km of mainland China coast between 18°N–23°N and east of 109°E. Season considered is May–November (covering 99% of events). Only TCs with lifetime maximum intensity >13 m s−1 were included. Basin Accumulated Cyclone Energy (ACE) computed as sum over 3-hourly periods of squared maximum sustained wind, excluding extratropical phases.
• Oceanic and atmospheric fields: Subsurface ocean temperatures from Met Office Hadley Centre EN4 (with time-varying bias adjustments). Alternative subsurface products: ECMWF ORAS5 and NOAA GODAS. SST from HadISST. Thermocline depth estimated by linear interpolation to the 20 °C isotherm. Atmospheric reanalysis fields (winds, geopotential height, humidity, vertical wind shear) from ERA5, monthly means on 1°×1° grids.
• Predictor definition: subNiño4 = June–August (JJA-1) mean subsurface temperature anomaly averaged over the equatorial central Pacific warm pool region analogous laterally to Niño 4 and vertically between 100–300 m (near thermocline/undercurrent depths). Correlation maps were computed to identify regions/depths of strongest relationship to next-year South China landfall counts and to verify isolation from surface in prior summer.
• Statistical modeling: Compared Poisson regression (typical for count data) with linear regression for predicting seasonal landfall counts using subNiño4 as predictor. Model evaluation used leave-one-out cross-validation (LOOCV). Additionally, a double cross-validation was implemented: an outer LOOCV for model validation and an inner loop to select the optimal predictor box from an ensemble (1755 candidates) spanning positions (146°E–224°E, 11°S–11°N at 2° lon × 1° lat spacing), depths (50–350 m in 50 m steps), and months (May–September), with box size fixed to 5° lon × 10° lat and 200 m depth over 3 months. This guards against bias from predictor selection on the full dataset.
• Skill metric: Skill against climatology computed as Skill = (1 – RMSE/RMSE_c) × 100%, where RMSE_c is the RMSE of climatological-mean predictions.

• Equatorial central Pacific subsurface temperature anomalies at 100–300 m during prior summer (JJA-1) are strongly anti-correlated with the following year’s South China TC landfall counts; reported correlation magnitude r = 0.67 (p < 0.001). Cold (warm) subsurface anomalies precede higher (lower) landfall counts.
• The prior-summer Pacific SST shows no significant correlation with the following year’s South China landfalls, emphasizing the subsurface origin of predictability at year lead.
• The depth of the equatorial thermocline near the dateline also predicts next-year landfall (r = 0.63, p < 0.001), though slightly weaker than subNiño4.
• Relationship to basin metrics: prior-summer subNiño4 correlates with basin ACE (r = 0.45, p = 0.006) but has weak correlation with total NW Pacific TC number (r = −0.26, p = 0.12). Landfall count vs. basin storm total is weak (r = 0.28, p = 0.09). Basin ACE and South China landfall are anti-correlated (r = −0.40, p = 0.01).
• Spatial signals: Low subNiño4 years are followed by La Niña–like SST patterns (central Pacific cooling, warmer SSTs east of Philippines) with very high spatial correlation to Niño 3.4 SST anomaly fields. The dominant, regionally consistent atmospheric signal linked to increased South China landfall is enhanced northwestward steering at ~500 hPa (zonal wind/geopotential height anomalies) over the South China Sea and Philippine Sea.
• Other environmental fields (SST, vertical wind shear, low-level vorticity, mid-level humidity) show mixed/favoring-opposing patterns across the South China and Philippine Seas, reinforcing steering as the primary consistent driver.
• Genesis location shifts toward South China in low subNiño4 years are small and not statistically significant (p > 0.1).
• Predictive performance: A linear regression using subNiño4 achieves cross-validated r = 0.61 (p < 0.001) and outperforms Poisson regression in correlation, RMSE, and skill. Double cross-validation that re-selects the predictor each fold yields an optimal box shifted ~8°E and 4°N of the Niño 4 box (still 100–300 m, JJA-1), with double cross-validated r = 0.65 (p < 0.001) and forecast skill of 23% versus climatology.
• Comparisons to ENSO indices: Prior-summer subNiño4 correlates with the subsequent TC-season Niño 3.4 (r = 0.53, p < 0.001). In-season Niño 3.4 correlates with landfall at r = −0.46 (p = 0.004), weaker than subNiño4’s year-ahead relationship.

The study proposes that subsurface heat content anomalies in the equatorial central Pacific (subNiño4) precede and help set the subsequent year’s surface ENSO state by about 12 months via recharge–discharge processes, advection, and upwelling. When these anomalies emerge at the surface, they alter tropical diabatic heating patterns, leading to large-scale atmospheric circulation responses (Matsuno–Gill–type), including strengthened northwestward mid-tropospheric steering winds that favor TC landfalls along the South China coast. The stronger year-ahead predictive relationship from subNiño4 compared to in-season Niño 3.4 is attributed to radiative-cloud feedbacks that damp in-season SST anomalies (reducing their direct correlation with landfall), while the subsurface index remains an independent, leading indicator unaffected by contemporaneous TC-related cloudiness. Thus, subsurface ocean thermal structure provides a physically grounded and more temporally leading predictor of regional TC landfall risk than traditional surface ENSO indices alone.

The authors define a subNiño4 index as the JJA-1 subsurface temperature anomaly averaged over the equatorial central Pacific warm pool at 100–300 m depth and show it is strongly anti-correlated with the following year’s South China TC landfall count. They demonstrate a simple, robust predictive model that can be issued nearly a year in advance with a double cross-validated skill of 23% against climatology, comparable to some spring-issued forecasts. The mechanism links subsurface anomalies to subsequent ENSO SST patterns and associated atmospheric steering changes that increase South China landfall likelihood in La Niña–like conditions. The work suggests that leveraging regional subsurface ocean temperature could extend the lead time of forecasts for other atmospheric variables and regions beyond current expectations.

• The reported correlation between subNiño4 and landfall counts is sensitive to the lifetime maximum intensity threshold used to select TCs making landfall.
• While steering anomalies are regionally consistent, other environmental factors (SST, vertical wind shear, low-level vorticity, mid-level humidity) show heterogeneous and sometimes opposing signals across the South China and Philippine Seas, complicating attribution.
• Shifts in genesis location for landfalling storms are small and statistically insignificant, indicating that predictability primarily reflects steering rather than genesis changes.
• Although robustness was checked across three subsurface ocean datasets (EN4, ORAS5, GODAS), results remain contingent on data quality and reanalysis uncertainties.
• The optimal predictor region identified via double cross-validation differs slightly from the standard Niño 4 box, indicating some sensitivity to predictor definition and emphasizing the need for careful out-of-sample selection procedures.