OceanLSTM: xLSTM with spatial attention for salt spray formation and migration prediction in marine hot-humid environments

Chuan Chen, Ganxin Jie, Jun Wang, S. Wang, Xiaonong XuFrontiers in Marine Science 2025 · DOI

Corrosion is the bane of marine engineering. In hot, humid environments, salt spray acts as a relentless "silent killer" of equipment and infrastructure. While we know that factors like wind speed and humidity drive salt spray formation, predicting exactly where and when it will strike has remained a complex challenge.

A new paper titled "OceanLSTM: xLSTM with spatial attention for salt spray formation and migration prediction in marine hot-humid environments" introduces a sophisticated deep-learning solution to this problem. By combining the latest advancements in sequence modeling with spatial intelligence, the authors have developed a tool that is both more accurate and significantly faster than current state-of-the-art methods.

TL;DR

OceanLSTM is a novel deep learning model that integrates Extended Long Short-Term Memory (xLSTM) with a spatial attention mechanism to predict salt spray migration. It achieves over 97% accuracy on major marine datasets while being computationally lighter and faster than previous models, making it a viable candidate for real-time corrosion monitoring.

What Problem Does This Paper Address?

Predicting salt spray isn't just about looking at a single sensor; it’s about understanding a dynamic "dance" between variables. Salt particles form, loft into the air, and migrate across geographic regions based on complex atmospheric conditions.

Traditional approaches fall into two camps:

  1. Physics-based simulations: These are accurate in controlled settings but are computationally "heavy" and struggle to adapt to the chaotic, large-scale data of the open ocean in real-time.
  2. Standard Machine Learning (RNNs/LSTMs): While better at handling data, standard LSTMs often suffer from "vanishing gradients" (forgetting long-term patterns) and struggle to account for the spatial relationships—how weather in one zone affects salt concentration in another.

The authors aim to bridge this gap by creating a model that understands both when things happen (temporal) and where they matter most (spatial).

Methods

The core of the paper is the OceanLSTM architecture, which rests on three innovative pillars:

  • The xLSTM Backbone: Unlike standard LSTMs, the "Extended" LSTM uses exponential gating and matrix-based memory cells. This allows the model to stabilize the learning process and retain critical environmental information over much longer time horizons without the math "breaking down" (vanishing/exploding gradients).
  • Spatial Attention Mechanism: Salt spray doesn't stay put. The model uses an attention layer to dynamically assign "weights" to different geographic regions. If a coastal zone is experiencing high winds, the model "pays more attention" to that area and its immediate neighbors to predict where the spray will migrate next.
  • Environmental Integration Strategy (EIS): To handle the messy reality of sensor data, the authors implemented dynamic feature scaling (to balance variables like temperature vs. salinity) and multi-resolution temporal smoothing to capture both sudden gusts and long-term seasonal trends.

Datasets

The model was put through its paces using four distinct datasets to ensure it could generalize across different environments:

  • NOAA Dataset: Long-term atmospheric data from oceanic stations worldwide (Open Ocean).
  • Marine Aerosol Dataset: Focused on coastal environments where wave-breaking creates high salt concentrations.
  • ASTM B117: Data from standardized laboratory salt spray tests (used as a controlled baseline).
  • ARGO Dataset: Global oceanic measurements including deep-sea temperature and salinity.

Key Insights and Ideas

  • Spatial Context is King: The "aha moment" of the study is that treating all geographic zones as equal dilutes predictive power. By adding spatial attention, the model mimics how a meteorologist looks at a map—focusing on the "source" regions of the spray.
  • Efficiency via Architecture: Surprisingly, OceanLSTM is more efficient than its predecessors. By using xLSTM’s refined gating, the model requires fewer parameters to achieve better results. It’s a "work smarter, not harder" approach to neural network design.
  • The "Homogeneity" Limit: Interestingly, the model didn't show massive gains on the ASTM B117 (lab) dataset. This confirms that the model’s strength lies in handling complexity; in a simple, uniform lab environment, the advanced spatiotemporal features are overkill.

Key Results

OceanLSTM consistently outperformed existing baselines (like standard LSTMs and ST-GCNs) across the board:

  • Accuracy: Reached 97.57% on the NOAA dataset and 98.20% on the Marine Aerosol dataset.
  • Computational Efficiency: The model uses roughly 177M parameters, compared to over 300M in some baseline models.
  • Speed: Inference time was clocked at approximately 106ms (NOAA), making it twice as fast as many competing architectures and well within the requirements for real-time alerting systems.
  • Ablation Success: Removing the spatial attention or the exponential gating led to dramatic drops in performance (accuracy falling by 5-10%), proving that these specific innovations are doing the heavy lifting.

Limitations and Future Work

While OceanLSTM is a major step forward, the authors note a few areas for improvement:

  • Anomalous Data: The model occasionally exhibits bias when dealing with extreme, "one-off" weather events that fall far outside the training distribution.
  • Multimodal Expansion: Future versions could benefit from integrating satellite imagery or oceanic current simulations directly into the feature set.
  • Uncertainty Quantification: The current model provides a "point prediction." The authors suggest that incorporating Bayesian neural networks or Monte Carlo Dropout could help provide "confidence intervals," which are vital for risk-averse marine engineering decisions.

For engineers and researchers in marine monitoring, OceanLSTM represents a shift toward more practical, deployable AI that doesn't sacrifice precision for speed.

Summary updated

Abstract

Introduction Salt spray formation and migration in hot and humid marine environments have a significant impact on marine engineering and equipment maintenance. Accurately predicting these phenomena is crucial for reducing corrosion damage. Traditional research methodologies primarily utilize statistical models or physics-based simulations. Although these approaches yield satisfactory results within controlled conditions, they often encounter limitations in accurately capturing the complexity and variability inherent to marine environments. These methods struggle to capture the spatiotemporal dependencies of salt spray formation and migration. Moreover, they are typically difficult to apply in real-time and lack the ability to handle large-scale, dynamic data. Methods This study aims to address this issue by proposing the OceanLSTM model, which combines the temporal modeling capabilities of xLSTM with a spatial attention mechanism to capture the spatiotemporal relationships between complex environmental variables, thereby improving the accuracy of salt spray predictions. Results The experiments used several representative marine environment datasets, including the NOAA and Marine Aerosol datasets. The experimental results demonstrate that OceanLSTM significantly outperforms traditional models in evaluation metrics such as accuracy and F1-score, especially on datasets with strong spatiotemporal dependencies. Discussion This research provides a more precise and efficient tool for future marine environment monitoring and corrosion prediction, offering important practical applications.
BibTeX
@inproceedings{chen2025oceanlstm,
  author = {Chuan Chen and Ganxin Jie and Jun Wang and S. Wang and Xiaonong Xu},
  title = {{OceanLSTM: xLSTM with spatial attention for salt spray formation and migration prediction in marine hot-humid environments}},
  booktitle = {Frontiers in Marine Science},
  year = {2025},
  doi = {10.3389/fmars.2025.1518050},
}
Figure