Congratulations to Anna Ćulibrk on the successful completion of her MSc thesis!
We are delighted to celebrate the successful completion of Anna’s MSc thesis, which explored the use of UAV imagery and deep learning semantic segmentation for the high-resolution seabed mapping at Pamfila Beach, Lesvos, Greece.
As one of the Mediterranean’s most valuable and threatened marine ecosystems, Posidonia oceanica plays a vital role in biodiversity, carbon sequestration, and coastal protection. Accurate habitat mapping is therefore essential for effective conservation and sustainable coastal management.
Anna developed and evaluated a DeepLabV3+ semantic segmentation workflow using MobileNetV2 and ResNet50 backbone architectures to classify four benthic habitat classes from UAV multispectral imagery.
📌 Key findings:
✅ Best-performing model achieved 90.12% Overall Accuracy, 0.84 Kappa, 0.75 mean IoU, and 0.86 Dice score.
✅ Increasing image tile overlap from 0% to 85% improved classification accuracy from 71.67% to 90.12%, demonstrating that tile configuration is a critical parameter in deep learning-based remote sensing workflows.
✅ MobileNetV2 significantly outperformed ResNet50, proving to be well suited for UAV-based seabed habitat mapping with limited training data.
✅ The proposed workflow provides a cost-effective, high-resolution solution for benthic habitat mapping, supporting the conservation of Posidonia oceanica and complementing existing satellite- and field-based monitoring approaches.
💻 Open Science
To encourage reproducible research and further development, the python code and the trained model are openly available on GitHub:
🔗 https://lnkd.in/dkNSi-bW
We congratulate Anna on this excellent achievement and wish her continued success in her future research and professional career. We are proud to support research that combines Remote Sensing, Artificial Intelligence, and GIScience to address real-world environmental challenges.
