Home/Industry Projects

Beneath the Waves: Deep Learning AI for Marine Habitat Mapping

A Deakin University study combining marine science and artificial intelligence demonstrates how deep learning can transform marine habitat mapping.

Challenge

Understanding the seafloor is vital for marine conservation, yet mapping benthic habitats remains a complex challenge. A new study by researchers from Deakin University’s Applied Artificial Intelligence Initiative and the Marine Mapping Lab shows how deep learning can dramatically improve this process.

Impact

Published in Remote Sensing in Ecology and Conservation (Q1), the paper “Comparing convolutional neural network and random forest for benthic habitat mapping in Apollo Marine Park” demonstrates the power of Convolutional Neural Networks (CNNs) to classify benthic habitats using multibeam bathymetry data. Compared to traditional machine learning methods like random forest, CNNs deliver significantly higher accuracy.

Research Partners

  • • Marine Mapping Lab – Deakin University

Awards

  • • Funded under Our Marine Parks Grants (4-HAZCNTR) awarded to Deakin’s Marine Mapping Lab.

Deep Learning for Benthic Habitat Mapping

The integration of artificial intelligence (AI) and remote sensing in marine conservation reflects a global shift toward scalable, data-driven solutions for marine spatial planning and biodiversity protection. As oceans face mounting pressures from climate change, overfishing, and habitat degradation, accurate and efficient mapping of benthic habitats has become critical.

Benthic habitats, the seafloor ecosystems that support biodiversity and provide essential services such as carbon sequestration, water filtration, and coastal protection, are fundamental to marine health. Understanding their distribution informs conservation strategies and resource management. However, mapping these habitats at high resolution is challenging due to the vastness and complexity of the seafloor.

This project leverages multibeam bathymetry and backscatter data, combined with advanced AI models, to improve habitat classification. Traditional approaches using random forest (RF) models rely on manually derived bathymetric variables, which can introduce subjectivity and limit scalability. In contrast, Convolutional Neural Networks (CNNs)automatically extract features from raw data, reducing manual intervention and potentially improving accuracy.

The study focuses on the northern section of Apollo Marine Park, a biodiversity hotspot with complex seafloor structures and transitional habitats that are difficult to classify. By comparing CNN and RF models, the research aims to:

  • Evaluate performance of CNNs versus RF for benthic habitat mapping.
  • Assess scalability and objectivity of deep learning approaches.
  • Advance automated mapping workflows for marine spatial planning.

Ultimately, this work seeks to establish best practices for integrating deep learning into marine habitat classification, supporting more accurate and efficient conservation strategies worldwide.

Solution

The study introduces a comparative approach to benthic habitat mapping using two machine learning models:

  1. Random Forest (RF) Model
    • Utilizes manually derived bathymetric variables (e.g., depth, slope, rugosity), wave height data, and spatial coordinates.
    • Represents the traditional method widely used in marine habitat classification.
  2. Convolutional Neural Network (CNN) Model
    • Trained directly on raw multibeam bathymetry data, combined with wave height and spatial information.
    • Eliminates the need for manual feature engineering by automatically learning spatial and textural patterns from the data.

By applying these models to high-resolution datasets from Apollo Marine Park, the authors aimed to determine whether CNNs could outperform RF models in accuracy and scalability. The CNN approach leverages its ability to capture multiscale spatial context, making it particularly effective for complex or transitional habitats where boundaries are less distinct.

Key Innovation:
The CNN model reduces subjectivity and complexity by bypassing manual derivation of bathymetric features, offering a more automated and potentially generalizable solution for large-scale marine habitat mapping.

Read the paper