Introduction
Motivation
Typhoons cause devastating damage, making accurate prediction critical. Traditional Numerical Weather Prediction (NWP) models are physics-based and reliable, but highly sensitive to uncertainty in initial conditions and observational gaps over ocean regions. Real observational data frequently contain noise and missing values, hindering the construction of accurate initial states. Data Assimilation (DA) addresses this by combining observations with model background states to produce more reliable initial conditions. Recent AI advances, particularly LSTM, have demonstrated strong capability in learning temporal patterns from sequential data.
Problem
Observational gaps in typhoon data are typically handled using simple linear extrapolation, which fails to capture the highly non-linear and dynamic nature of real typhoon movements. This distorted data directly degrades the accuracy of the DA system and the final forecast. An AI-based imputation approach is therefore needed to construct more reliable initial conditions. In particular, end-track missing data is the most common scenario in real maritime environments, where tracking systems lose coverage as typhoons move beyond observation network boundaries.
Goal
This study compares linear extrapolation and LSTM-based imputation for missing segments in Typhoon MAYSAK (2020) data — latitude, longitude, pressure, and wind speed. Each reconstructed observation dataset is then assimilated using EnKF with the same controlled synthetic background field. By evaluating the resulting tracks via RMSE, we assess how the choice of imputation method affects forecasting accuracy and demonstrate the potential of AI-driven data refinement.
Limitations of NWP
Physics-based models struggle with observational gaps over open oceans, leading to unreliable initial conditions.
Missing Observation Problem
Real typhoon data frequently contains missing values and noise, distorting the input to forecast systems.
Potential of AI-based Imputation
LSTM models excel at learning non-linear temporal patterns from sequential data — making them ideal for imputation.
End-track Data Loss
A Realistic Observation Gap in Maritime Monitoring
⚙️
Common in practice
Observation gaps frequently occur near the end of available records.
⚠️
Coverage limitations
Maritime environments often have sparse and inconsistent observations.
🔎
Operational impact
End-track gaps create a difficult and realistic forecasting challenge.
In real-world typhoon monitoring, observational gaps most commonly occur at the end of available data records. This is particularly evident in maritime environments, where tracking systems such as GPS, weather buoys, and reconnaissance aircraft have limited operational range and coverage. As a typhoon moves further into open ocean or beyond observation network boundaries, continuous real-time tracking becomes increasingly difficult, resulting in consecutive missing values at the tail end of the track record.
Therefore, end-track missing data represents the most operationally relevant and challenging scenario for typhoon forecasting systems.
key Terms
① Data Assimilation (DA)
A technique that combines observations with a background model state to estimate a more accurate atmospheric state
— ultimately improving the quality of input data for weather forecasting.
② LSTM (Long Short-Term Memory)
An AI model that learns temporal patterns from sequential data. Unlike simple linear methods, LSTM can learn non-linear temporal patterns from sequential data, making it useful for imputing missing typhoon track values — making it powerful for predicting and imputing missing values in typhoon track data.
Research Question
Under the same controlled background field, can LSTM-based imputation produce more accurate EnKF-based typhoon track forecasts than linear extrapolation?
Controlled Condition
The same synthetic background field is used in both cases to isolate the effect of the imputation method.
Comparison
Case A uses linear extrapolation, while Case B uses LSTM-based imputation for the missing end-track observations.
Hypothesis
If LSTM captures non-linear typhoon movement more effectively, Case B will show lower RMSE after EnKF assimilation.