Typhoon Center Estimation

Updated 6 April 2026

Typhoon center estimation is the rigorous identification or prediction of a storm’s central coordinates, fundamental for accurate tracking and forecasting.
Methodologies include image-based keypoint localization, object detection, and hybrid physics-ML models that balance precision, efficiency, and uncertainty quantification.
Practical applications span real-time operational detection, advanced numerical weather prediction initialization, and enhanced hazard modeling in meteorological regimes.

Typhoon center estimation refers to the rigorous identification or prediction of the instantaneous geographic coordinates (latitude, longitude) of the center of a typhoon’s circulation—variously called the “typhoon center,” “eye,” or “surface circulation center.” Accurate center estimation is foundational to operational typhoon detection, real-time tracking, intensity analysis, numerical weather prediction (NWP) initialization, assimilation pipelines, and downstream hazard modeling. Contemporary research in typhoon center estimation encompasses observational algorithms, statistical and deterministic dynamical models, advanced deep learning frameworks, and hybrid physical-ML architectures, each optimizing for precision, uncertainty quantification, or operational efficiency in a variety of input modalities and meteorological regimes.

1. Problem Formulation and Core Methodological Variants

Typhoon center estimation naturally splits into two primary subdomains: (a) operational “center-fixing” for existing storms (i.e., per-frame estimation from observations), and (b) multi-step trajectory or track prediction (short- to long-term forecasting). In both cases, the target output is a sequence $\{\hat{y}_t\}_{t=1}^T$ with each $\hat{y}_t = (\phi_t, \lambda_t)$ , interpreted as the center location in geodetic coordinates. Inputs span multi-channel satellite imagery (infrared, visible, microwave), scalar meteorological time series (e.g., wind, pressure), gridded reanalysis products, and increasingly, auxiliary semantic representations (LLM-generated meteorological descriptions).

For center-fixing, methods structurally include:

Image-based keypoint localization: Inputs are 2D or 3D (spatiotemporal) satellite image tensors; outputs are coordinates or probabilistic maps highlighting the center (“heatmap regression”).
Object-detection formulations: Treated as a degenerate object-detection task—identifying a single (x, y) point per frame with/without explicit bounding box prediction (Tan, 2020, Kitamoto et al., 2024).
Physical/heuristic algorithms: Based on extrema in sea-level pressure, vorticity, or pattern-matching in the image domain (Niu et al., 2024, Lagerquist et al., 2024).

For forecasting, methods fit into:

Statistical/stochastic trajectory generation: Markovian or AR models separately estimate bearing and speed increments, with lysis modeled by logistic or empirical distributions (Dalmasso et al., 2019). Ensembles produce “cones of uncertainty.”
Deep learning time series regressors: Transformers, LSTMs, or diffusive models directly regress future center tracks from recent histories of multivariate fields (Li et al., 21 Jun 2025, Park et al., 2024).
Physics-aware and hybrid ML-dynamical models: Physics-conditioned encoders, spectral nudging, and data assimilation combine physically plausible reconstructions with ML’s ability to correct biases or subgrid variability (Park et al., 2024, Niu et al., 2024).

2. Observational Center Estimation: Algorithms and Benchmarks

Satellite Image-Based Approaches

Image-based center-fixing employs deep convolutional architectures trained with point-annotation or heatmap regression paradigms. TCLNet (Tan, 2020) exemplifies a fully convolutional encoder–decoder for IR images, outputting Gaussian heatmaps centered on meteorologist-provided eye labels. For a 512×512 IR image, TCLNet produces a 128×128 heatmap, with the predicted center at the global maximum. The network is optimized with a “piecewise” TCL+ loss, which down-weights noisy or ambiguous samples (notably “non-eyed” storms). On the TCLD dataset (Huanjing-8 IR imagery), TCLNet achieves a mean localization error (MLE) of 4.51 px over all samples, outperforming deeper baselines while being 92.7% more parameter-efficient.

The Digital Typhoon V2's center estimation task (Kitamoto et al., 2024) uses a U-Net trained with a weighted Hausdorff loss for object-detection–style centering on 256×256 IR crops. Performance is strongly grade-dependent: mean center errors drop to 15 km for grade-5 (clear-eyed) storms, but rise to 65 km for grade-2 (weak, disorganized) systems. Subtle challenges—such as cross-hemispheric generalization and augmentation-induced artifacts—are addressed through basin-specific horizontal flipping and conservative crop/rotation policies.

GeoCenter (Lagerquist et al., 2024) extends image-based center-fixing to high-cadence (10-min) geostationary IR time series, leveraging temporal context across 10 IR channels and correcting initial “first guess” offsets. GeoCenter outputs not just a deterministic fix but a calibrated ensemble, yielding mean errors of 25–27 km (all systems), with gradated performance for weaker vs. intense systems.

Heuristic and Hybrid Methods

Operational center-fixing via physical fields typically exploits the collocation of the SLP minimum and cyclonic vorticity maximum. The Pangu_SP hybrid model (Niu et al., 2024) combines ML-derived synoptic fields (Pangu-Weather) with high-resolution WRF, searching a 10°×10° window for the SLP or 850–950 hPa vorticity extremum, with sub-grid quadratic interpolation for enhanced precision. Ensemble and smoothing post-processing enforce dynamically plausible evolution.

3. Deep Learning and Physics-Conditioned Forecasting

Transformer Architectures with Language Augmentation

TyphoFormer (Li et al., 21 Jun 2025) formulates track forecasting as auto-regressive sequence modeling of future centers, conditioning on both numerical time series from HURDAT2 and LLM-generated textual meteorological prompts. Numerical and language embeddings are fused at each time step via a prompt-gated fusion (PGF) mechanism, followed by L-layer Transformer encoding. The decoder auto-regressively propagates center coordinates; training minimizes MSE in latitude/longitude.

When evaluated on HURDAT2 for 24-h lead forecasts, TyphoFormer reduces the mean absolute error to 0.312 normalized units (ΔR=49.56 km), surpassing both TSMixer and classical CLIPER baselines. The explicit integration of LLM prompts provides high-level semantic cues (e.g., quadrant wind radii, evolution summaries) that bolster the model’s performance—particularly in sparse-historical or nonlinear deviation regimes.

Hybrid ML–Physics Systems and Bias Correction

The LT3P system (Park et al., 2024) provides physics-conditioned long-term trajectory prediction without requiring reanalysis at forecast time. It comprises:

Physics encoder $R(\cdot)$ : pretrained on ERA5 to capture spatiotemporal features from fundamental fields (geopotential, u, v) across multiple pressure levels.
Bias corrector $B(\cdot)$ : transforms real-time NWP (UM) fields into the ERA5 domain for transferability.
Trajectory regressor $D(\cdot)$ : cross-attends corrected physics features with past observed centers, directly regressing future geodetic coordinates.

The LT3P pipeline achieves a 72-h final displacement error (FDE) of 70.9 km (ERA5 inputs) and 143.0 km (real-time UM inputs), outperforming ensemble NWP and prior ML models. For applications demanding real-time tractability and minimized reanalysis latency, this approach offers a quantifiable advance.

Spectral nudging—used in Pangu_SP (Niu et al., 2024)—injects large-scale ML forecast information into the regional WRF simulation, constraining synoptic behavior while preserving WRF’s capacity for mesoscale intensification. Post-hoc assimilation of water-vapor radiances (FY-4B AGRI) further sharpens initialization, with demonstrated reductions in mean absolute track error versus standard NWP initializations.

4. Statistical, Stochastic, and Uncertainty Quantification Approaches

Many operational workflows require not just point estimates, but full trajectory ensembles or uncertainty quantification on future positions. The flexible pipeline by Hosseini et al. (Dalmasso et al., 2019) instantiates such inference entirely from official HURDAT2 best-track histories via:

Data-driven, block-wise (10°×10°) linear or autoregressive (AR) models for storm bearing and speed increments.
Lysis (death) modeled via logistic regression or kernel-smoothed life-span density.
Ensemble generation by sequential sampling (M≈350 replicates).

Prediction bands (“cones of uncertainty”) are then generated via:

Pointwise kernel-density or spherical bands (analogous to operational NOAA cones but ensemble-based).
Uniform-in-time convex-hulls or δ-ball methods, optimizing for pathwise or timewise coverage.
Empirical calibration based on withheld tracks enables coverage control and band width adjustment.

Such frameworks support modular substitution of ML predictors or innovation in depth-based band construction, allowing end-to-end, transparent uncertainty quantification.

Uncertainty quantification is operationalized in GeoCenter (Lagerquist et al., 2024) using deep ensemble techniques, CRPS-based training, and diagnostic calibration (spread–skill reliability, rank-histogram deviation, monotonic discard tests). Forecast-user feedback and agile response to confidence intervals (e.g., low-confidence samples flagged for analyst review) are thus easily embedded.

5. Classical Analytical and Physical Modeling Foundations

Mathematical models grounded in primitive equations, such as those by Rozanova et al. (Rozanova et al., 2010), provide analytic insight into center trajectory behaviors. In the $l$ -plane approximation, where the Coriolis parameter is treated as fixed, the eye’s trajectory admits explicit solutions as superpositions of two circles (periods $2\pi/l$ , $2\pi/b_0$ ; $l$ is the local Coriolis parameter, $b_0$ is the vortex-center vorticity). The model parameters are fitted in real time via an inversion algorithm using three consecutive fixes. During the mature (conservative) vortex phase, this analytic scheme approximates observed center trajectories with position errors of 20–50 km over 2–3 days, capturing complex structures such as loops or recurve points.

Limitations emerge from neglected effects: ambient steering flows, β-drift, surface roughness, and transitions out of the steady-vortex phase. The parameter-adopting approach remains relevant as a physically interpretable benchmark and for rapid analytic forecast in operational workflows.

6. Evaluation Metrics, Datasets, and Performance Benchmarks

Quantitative evaluation of center estimation leverages several key metrics, contingent on the nature of the input and prediction horizon:

Metric	Definition / Use
Great-circle distance E	$\hat{y}_t = (\phi_t, \lambda_t)$ 0 [km]
Mean/Median/RMS Error	Used universally for direct center estimations (Lagerquist et al., 2024)
Mean localization error	Pixel/physical distance between predicted and true center
FDE, ADE (Forecast)	Final/Average displacement error over multi-step track forecasts (Park et al., 2024)
Hausdorff Distance	Weighted Hausdorff loss for keypoint/heatmap fit (Kitamoto et al., 2024)
CRPS	Calibrated ensemble sharpness-spread metric (Lagerquist et al., 2024)

Gold-standard datasets include HURDAT2 (NOAA), Digital Typhoon V2, PHYSICS TRACK, and Huanjing-8(Satellite) TCLD. Graded error reporting by storm strength, intensity, or hemisphere enhances interpretability. For example, U-Net in (Kitamoto et al., 2024) achieves ≈15 km error for grade-5 eyes and ≈65 km for grade-2 storms. GeoCenter achieves 14.6 km RMS error for category 2–5 hurricanes, matching high-latency microwave-based fixes (Lagerquist et al., 2024). State-of-the-art deep learning track forecasters such as TyphoFormer and LT3P yield <50–70 km error at 24–72 h horizons (Li et al., 21 Jun 2025, Park et al., 2024).

7. Current Challenges and Prospective Directions

Challenges in typhoon center estimation reflect both data-specific and generalizable concerns:

Weak/disorganized systems: Estimation error remains high where center signatures are poorly defined, e.g., in grade-2 or non-eyed TCs (Tan, 2020, Kitamoto et al., 2024, Lagerquist et al., 2024).
Cross-basin generalization: Morphological and circulation differences require preprocessing (e.g., hemisphere-specific image flipping) and careful basin-specific model tuning (Kitamoto et al., 2024).
Physics–ML integration: Emerging methods exploit spectral nudging, data assimilation, and simplified physical constraints to bridge accuracy and interpretability gaps (Niu et al., 2024, Park et al., 2024), but generalization to all basins and coupling with operational workflows require further study.
Uncertainty characterization: Reliable, calibrated ensemble prediction and confidence quantification for center-fix and track estimates are essential for risk management (Dalmasso et al., 2019, Lagerquist et al., 2024).
Semantic augmentation: LLM-generated prompts can supply high-level meteorological priors, improving prediction for sparse or nonlinear tracks at the cost of potential prompt-induced biases (Li et al., 21 Jun 2025).

Future research is focused on integrating additional observational modalities (scatterometer, water-vapor, visible imagery), further automating uncertainty quantification, cross-task multi-objective training (joint center and intensity estimation), and operational deployments with strict latency and reliability constraints.

Overall, modern typhoon center estimation encapsulates a spectrum from physics- and heuristic-based algorithms to cutting-edge deep learning approaches leveraging semantically enriched input representations, with steady progress in operational accuracy, robustness, and uncertainty management across both observational and forecasting domains.