Arman NeyestaniComputer Vision · MeasurementResume
← All work

Environmental Sensing

Significant wave-height forecasting

A GRU model turns 16-step marine observations into a one-step-ahead significant-wave-height estimate for digital-twin prototypes.

Inference repository ↗IEEE paper ↗
My role
I led the study as first author, developing preprocessing, data splits, multivariate GRU modeling, and evaluation with my co-authors.
What I built
A three-layer GRU predicts the next VHM0 value. The repository packages the trained model, fitted scalers, sample data, validation, and a deterministic inference interface.
Key results
  • 0.182 m best reported RMSE (Barcelona evaluation)
  • 0.9517 best reported Pearson r (Barcelona evaluation)
  • 0.0955 m MAE (Tarragona evaluation)
  • 0.208 m RMSE (Tarragona evaluation)
Stack
Python · TensorFlow · GRU · Time-series preprocessing · Saved scalers · Deterministic inference
Observed and predicted significant wave-height time series for the Barcelona evaluation.
Repository-held paper figure showing prediction behavior for the Barcelona evaluation.

Problem

What needed to be understood or measured

Marine monitoring systems need compact, reproducible inference that preserves the preprocessing used during model development.

My contribution

My contribution

I led the study as first author, developing preprocessing, data splits, multivariate GRU modeling, and evaluation with my co-authors.

Experimental setup

Data and setup

Four wave features over a 16-step look-back window, evaluated across Tarragona, Barcelona, and EMSO-OBSEA measurements.

OBSEA underwater observatory and buoy visualized in a digital twin interface.Open figure —
Publication figure

OBSEA observatory and buoy data represented in the marine digital-twin interface.

A Significant Wave Height Data-Driven Modeling for Digital Twins of Marine Environment, Fig. 1, p. 2.

Method

From input to interpretable output

A three-layer GRU predicts the next VHM0 value. The repository packages the trained model, fitted scalers, sample data, validation, and a deterministic inference interface.

Three-layer GRU architecture with preprocessing, dropout, dense prediction and Adam optimization.Open figure —
Publication figure

Published GRU architecture and 16-by-5 input sequence.

Significant Wave Height paper, Fig. 2, p. 2.

Results

Reported results

0.182 m

best reported RMSE

Barcelona evaluation

0.9517

best reported Pearson r

Barcelona evaluation

0.0955 m

MAE

Tarragona evaluation

0.208 m

RMSE

Tarragona evaluation

Actual and predicted Barcelona wave height with a 95 percent prediction interval.Open figure —
Publication figure

Barcelona actual and predicted VHM0 with the published 95% prediction interval.

Significant Wave Height paper, Fig. 4, p. 4.
Actual and predicted EMSO-OBSEA wave height with prediction interval.Open figure —
Publication figure

EMSO-OBSEA actual and predicted VHM0 with the published prediction interval.

Significant Wave Height paper, Fig. 7, p. 5.
Published cross-dataset evaluation
Tarragona
0.208 m
0.9354
Barcelona
0.182 m
0.9517
EMSO-OBSEA
0.252 m
0.8857

Values reported in the 2024 IEEE MetroSea paper. Bars use separate scales and are labeled directly.

Validation & uncertainty

Reference and uncertainty checks

Metrics are reported per measurement site in the paper. The repository provides inference assets and checks, but intentionally does not include a retraining pipeline.

Barcelona wave-height prediction time series with flagged outliers.Open figure —
Publication figure

Barcelona time series with residual-based outliers flagged.

Significant Wave Height paper, Fig. 5, p. 4.

Limitations & failure modes

Operating limits and failure modes

EMSO-OBSEA wave-height series with flagged outliers.Open figure —
Publication figure

EMSO-OBSEA prediction series with flagged outliers.

Significant Wave Height paper, Fig. 8, p. 5.

Distribution shift between sites, missing observations, sensor drift, extreme sea states, and preprocessing mismatch can reduce forecast quality.

The repository is inference-ready but intentionally does not include a retraining pipeline. Metrics shown here are paper results, not newly recomputed benchmarks.

Technical stack

Tools selected for the measurement chain

  • Python
  • TensorFlow
  • GRU
  • Time-series preprocessing
  • Saved scalers
  • Deterministic inference

Sources & project links

Papers and code

IEEE MetroSea 2024 paper and MIT-licensed public inference repository. Metrics shown here are paper results, not newly recomputed benchmarks.

Inference repository ↗IEEE paper ↗