⚡ electrical engineering

Significant Wave Height Prediction and Harbour Entry Risk Index (HERI) in the Sunda Strait using ERA5-Based Deep Learning to Support Merak- Bakauheni Ferry Operations

This study develops a CNN-LSTM and BiLSTM-based deep learning framework using ERA5 data to predict significant wave height in the Sunda Strait and introduces a novel Harbour Entry Risk Index (HERI) that significantly outperforms traditional wave height thresholds in detecting berthing risks, thereby enabling effective 6–12-hour early warnings for Merak-Bakauheni ferry operations.

Original authors: Muhammad A.R. Fauzi, Trika A. Tarigan, Ulil Amri

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad A.R. Fauzi, Trika A. Tarigan, Ulil Amri

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every year, millions of people in Indonesia travel across the Sunda Strait, the narrow body of water separating the islands of Java and Sumatra. This route is the busiest ferry crossing in the nation, serving as a vital artery for passengers, vehicles, and essential goods. During the massive annual holiday known as Lebaran, the number of travelers swells to nearly six million, creating a logistical challenge that the country calls its national barometer for smooth travel. For the ferry operators, safety is paramount, but the ocean is unpredictable. The primary danger comes from the waves. Current regulations dictate that ferries must stop crossing if the significant wave height—the average height of the highest third of the waves—reaches 2.5 meters. However, this simple rule has a blind spot. It assumes that the height of the wave is the only thing that matters, ignoring how the wave moves, how long it takes to pass, and the direction from which it strikes the ship. A wave that is not particularly tall can still be deadly if it hits the ferry at the wrong angle or if its rhythm matches the natural rocking motion of the vessel, causing it to roll dangerously.

To solve this problem, a team of researchers set out to create a smarter way to predict danger in the Sunda Strait. They focused on the specific stretch of water between the ports of Merak and Bakauheni, where the water is deep and the currents are complex. Instead of relying on a single number to decide if it is safe to sail, they developed a new tool called the Harbour Entry Risk Index. This index acts like a comprehensive health check for the sea, combining three different factors: the height of the waves, the timing of the waves relative to the ship's natural rocking speed, and the direction the waves are coming from. To make this tool work, the researchers needed to know what the waves would do hours in advance. They turned to artificial intelligence, specifically a type of computer learning designed to recognize patterns in time-based data. They trained several different computer models on fourteen years of historical ocean data, teaching them to look at wind speed, pressure, and past wave behavior to forecast what the waves would look like six, twelve, and twenty-four hours into the future.

The researchers tested five different types of computer models to see which one could predict the waves most accurately. They found that a hybrid model, which combined the ability to spot local patterns with the ability to remember long-term sequences, performed the best for short-term forecasts. This model could predict the wave height six hours ahead with remarkable precision, making errors of less than six centimeters. For longer forecasts, up to a full day ahead, a different type of model that processes information in both forward and backward directions proved slightly more reliable. These computer models were far superior to a simple method that assumes the waves will stay exactly the same as they are right now. The key breakthrough, however, was not just in predicting the height, but in how that prediction was used. When the researchers applied their new risk index to the historical data, they discovered a critical flaw in the old safety rules. The single threshold of 2.5 meters never triggered a warning at the main ferry lane during the entire fourteen-year study period, even though the ocean conditions were often rough. In fact, the highest waves recorded at that specific location never exceeded 2.4 meters.

This finding revealed that the old rule was failing to detect real danger. The researchers then tested their new risk index against a record of six actual berthing incidents that occurred at the Merak port between 2020 and 2024. Under the old system, the safety threshold would have remained silent for all six of these dangerous events because the waves were technically below the 2.5-meter limit. In contrast, the new risk index correctly identified all six incidents as dangerous. It did this by recognizing that the combination of moderate wave heights, specific wave periods that matched the ferry's rocking rhythm, and waves hitting the ship head-on created a hazardous situation that a simple height measurement would miss. The index also showed a clear seasonal pattern, with the highest risk occurring during the monsoon months of December and January, which aligns perfectly with the historical record of when accidents actually happened.

The study further demonstrated that this system could provide a reliable early warning. Using the best computer models, the system could predict dangerous conditions six to twelve hours before they arrived. This lead time is crucial for ferry operators, giving them enough notice to adjust schedules or delay departures before long lines of vehicles form on the roads leading to the port. The system was designed to be cautious; when it made a mistake, it almost always underestimated the danger slightly rather than overestimating it, ensuring that a truly dangerous hour was never missed entirely. While the researchers noted that their work relied on historical data and that future improvements could include real-time sensors and more detailed maps of the harbor, the results offer a powerful new way to manage safety. By moving beyond a single number and embracing a more complete picture of the ocean's behavior, this approach provides a practical, data-driven foundation for keeping the millions of travelers on the Sunda Strait safe.

Technical Summary: Significant Wave Height Prediction and Harbour Entry Risk Index (HERI) in the Sunda Strait

Problem Statement
The Merak–Bakauheni ferry route in the Sunda Strait is Indonesia's busiest inter-island crossing, handling approximately six million passengers annually during the Lebaran (Eid al-Fitr) season. Current safety regulations (KSOP) mandate crossing suspensions when Significant Wave Height (SWH) exceeds 2.5 m. However, historical analysis of the central ferry lane reveals that SWH at the primary operational grid point never exceeded 2.405 m between 2010 and 2024, meaning the regulatory threshold was never triggered. Despite this, six documented berthing incidents occurred at Merak Harbour between 2020 and 2024 under these "safe" wave heights. The study identifies that single-parameter SWH thresholds fail to account for critical multi-factor hazards: wave period resonance with the ferry fleet's natural roll period, and wave direction relative to the berth. Existing machine learning (ML) studies in Indonesia have focused on the Java coast, leaving the Sunda Strait without a dedicated deep learning-based prediction framework.

Methodology
The study employs a data-driven approach using ERA5 reanalysis data (2010–2024) at a 0.25° resolution, extracted from a single grid point (−6.0°S, 105.5°E) representing the center of the ferry lane.

SWH Prediction Models: Five architectures were trained and compared for three prediction horizons ( $t+6$ h, $t+12$ h, $t+24$ h):
- Deep Learning: Stacked LSTM, GRU, BiLSTM, and CNN-LSTM.
- Tree-Based: XGBoost.
- Baseline: A persistence model ( $SWH_{t+h} = SWH_t$ ).
- Data Configuration: The dataset was split chronologically (Training: 2010–2022; Validation: 2023; Test: 2024). Input features included SWH, wind speed/direction, mean wave period (MWP), mean wave direction (MWD, encoded as sin/cos), and mean sea level pressure. A 72-hour lookback window was used.
Harbour Entry Risk Index (HERI): A novel composite index was developed to quantify berthing risk. Unlike simple thresholds, HERI combines three parameters:
- Normalized SWH (anchored to the 2.5 m regulatory limit).
- A resonance factor ( $R$ ) based on the proximity of the mean wave period to the fleet's estimated natural roll period (5–11 s).
- A directional factor ( $F_{dir}$ ) based on the angle of wave approach relative to the berth.
- Formula: $HERI(t) = \text{clip}[ W_h \cdot SWH_{norm}(t) \cdot R(t) + W_d \cdot F_{dir}(t); 0, 1 ]$ .
- Calibration: The index was calibrated against six documented Merak incidents (2020–2024) to set the "Dangerous" threshold at 0.50. The authors explicitly note that these six events constitute a calibration set rather than an independent validation.
Evaluation: Model performance was assessed using RMSE, MAE, MAPE, and $R^2$ . The predictive HERI system was evaluated using confusion matrices, focusing on recall (true positive rate) and false negative rates (FNR) to ensure safety-critical detection.

Key Results

SWH Prediction Performance:
- CNN-LSTM achieved the best performance for short-to-medium horizons ( $t+6$ h: RMSE = 0.0579 m, $R^2$ = 0.9459; $t+12$ h: RMSE = 0.0892 m, $R^2$ = 0.8716). Its architecture effectively extracted local patterns via Conv1D layers before modeling temporal dependencies.
- BiLSTM excelled at the long horizon ( $t+24$ h: RMSE = 0.1307 m, $R^2$ = 0.7243), benefiting from bidirectional sequence processing.
- All deep learning models significantly outperformed the persistence baseline and XGBoost, which degraded faster at longer horizons.
HERI Detection Capability:
- The single 2.5 m SWH threshold detected 0% of the six documented Merak incidents.
- The HERI index detected 100% of the six specific incidents used for calibration (2020–2024), with values exceeding the "Dangerous" threshold (0.50) by margins of +0.05 to +0.20. The authors clarify that because these six events informed the threshold, they serve as a calibration set rather than an independent validation.
- The seasonal distribution of "Dangerous" HERI hours (peaking in December–January) aligned coherently with the monthly distribution of a broader 16-incident record (2015–2024), providing independent seasonal validation.
Predictive HERI Performance:
- Using predicted SWH inputs, the system achieved category accuracies of 93.9–97.3% and dangerous-condition recalls of 82.8–94.6% across horizons.
- Fail-Safe Mechanism: A critical finding was that no actual dangerous hour was ever predicted as "Safe" (skip $_2$ = 0). All false negatives were downgraded only one category (e.g., predicted "Caution" instead of "Dangerous"), ensuring an alert state is always triggered.
- The 6–12 hour lead time provided by the models is deemed sufficient for operational adjustments (e.g., schedule changes) before vehicle queues form.

Significance and Contributions
The paper claims three primary contributions:

Academic: It presents the first machine learning wave prediction study specifically targeting the Sunda Strait and introduces HERI as a novel multiparameter framework that moves beyond single-threshold SWH criteria.
Practical: The integration of CNN-LSTM/BiLSTM predictions with HERI classification offers a deployable 6–12 hour early-warning system compatible with the ASDP ferry operator's scheduling cycle.
Policy: The study provides a 14-year empirical evidence base and a verified incident dataset to support the transition toward multiparameter risk management in Indonesian ferry safety regulations.

Limitations and Future Work
The authors note that the six Merak incidents used for calibration constitute a calibration set rather than an independent validation. Furthermore, the Bakauheni side lacks positive validation cases as its incidents are attributed to currents rather than waves. The study also relied on expert judgment for weight calibration due to a lack of ASDP operational log data. Future work is recommended to empirically calibrate HERI weights, develop multi-output models (predicting SWH, MWP, and MWD simultaneously), and apply spatial downscaling using SWAN models for higher-resolution harbor basin analysis.

Technical Summary: Significant Wave Height Prediction and Harbour Entry Risk Index (HERI) in the Sunda Strait

More like this