The AXOLOTL project’s Research and Innovation (R&I) component tackles pressing challenges in marine biodiversity assessment and maritime surveillance by leveraging advanced deep learning (DL) methodologies. By addressing the complexities of in-situ and remote sensing data, AXOLOTL aims to develop robust AI-based solutions that enhance the detection, classification, and monitoring of marine species and vessels, enabling more efficient conservation and surveillance efforts.

 

 

Key Activities:

• Advanced ARMS Plate Image Analysis

Develop DL methodologies for analyzing highly heterogeneous image datasets from Autonomous Reef Monitoring Structures (ARMS), which are crucial for biodiversity assessment. The project addresses challenges such as inconsistent lighting, species labelling, and overlapping marine organisms by applying texture segmentation and anomaly detection techniques.

• Satellite Image-Based Vessel Detection

Address the limitations of medium-resolution satellite images (e.g., Sentinel-2) for small vessel detection using super-resolution techniques. AXOLOTL combines DL methodologies with auxiliary data sources like AIS and VMS to improve vessel detection accuracy while minimizing false positives. Domain adaptation techniques will also enhance the utility of super-resolved images.

• Underwater Acoustic Data Classification

Leverage DL to analyze underwater acoustic datasets for vessel detection and classification. This involves identifying specific acoustic signatures while addressing challenges like background noise and synchronization with auxiliary data sources. The project expands the network of hydrophones in Cyprus to enhance dataset availability.

• Integrated Data Preparation and Standardization

Standardize datasets from ARMS, satellite imagery, and underwater acoustic recordings into DL-friendly formats, ensuring compatibility for downstream applications. Collaborative efforts with project partners will ensure high-quality data processing and annotation.

• Real-World Validation of DL Models

Validate the developed DL methodologies across three use cases:

1. Species identification using ARMS images in Europe.
2. Small vessel detection from satellite images around Cyprus.
3. Acoustic-based vessel classification in the Belgian North Sea.

• Formulation of Methodological Hypotheses

Test research hypotheses, such as improving species labelling via unsupervised learning and enhancing vessel detection through super-resolution, to push the boundaries of DL applications in marine monitoring.