From underwater to aerial: a novel multi-scale knowledge distillation approach for coral reef monitoring

TL;DR

This paper presents a novel multi-scale knowledge-distillation framework for coral reef monitoring that fuses fine-scale underwater imagery with medium-scale aerial imagery. A transformer-based teacher model trained on underwater data guides a student aerial model through a footprint-weighted, soft-label distillation process, enabling high-resolution predictions over large reef areas. The approach achieves strong cross-scale alignment (AUC up to 0.9251 for annotations and 0.7952 for aerial predictions) and demonstrates a flexible, model-agnostic pathway to scale fine-grained coral classifications, with potential extension to satellite imagery and synchronized slow-moving species observations. This method advances scalable reef assessment and conservation by reducing annotation burden and enabling broad, accurate habitat mapping across expansive marine environments.

Abstract

Drone-based remote sensing combined with AI-driven methodologies has shown great potential for accurate mapping and monitoring of coral reef ecosystems. This study presents a novel multi-scale approach to coral reef monitoring, integrating fine-scale underwater imagery with medium-scale aerial imagery. Underwater images are captured using an Autonomous Surface Vehicle (ASV), while aerial images are acquired with an aerial drone. A transformer-based deep-learning model is trained on underwater images to detect the presence of 31 classes covering various coral morphotypes, associated fauna, and habitats. These predictions serve as annotations for training a second model applied to aerial images. The transfer of information across scales is achieved through a weighted footprint method that accounts for partial overlaps between underwater image footprints and aerial image tiles. The results show that the multi-scale methodology successfully extends fine-scale classification to larger reef areas, achieving a high degree of accuracy in predicting coral morphotypes and associated habitats. The method showed a strong alignment between underwater-derived annotations and ground truth data, reflected by an AUC (Area Under the Curve) score of 0.9251. This shows that the integration of underwater and aerial imagery, supported by deep-learning models, can facilitate scalable and accurate reef assessments. This study demonstrates the potential of combining multi-scale imaging and AI to facilitate the monitoring and conservation of coral reefs. Our approach leverages the strengths of underwater and aerial imagery, ensuring the precision of fine-scale analysis while extending it to cover a broader reef area.

Figures (13)

• Figure 1: On the left, an aerial orthophoto captures a complex coral assemblage. However, distinguishing between specific morphotypes is challenging because of the limited resolution. The red rectangle highlights the location of the fine-scale underwater image shown on the right. The underwater image provides a higher level of detail, allowing the identification of distinct coral morphotypes and specific classes, such as $\mathop{\mathrm{\textit{Algae}}}\nolimits$, that are difficult to identify in aerial imagery.
• Figure 2: Workflow of the multi-scale approach for coral reef monitoring.
• Figure 3: Visual georeferencing criteria to validate the georeferencing of underwater and aerial images with respect to the BD ORTHO® orthophoto. On the left (a) underwater images georeferenced with respect to the aerial orthophoto. On the right (b) aerial orthophoto georeferenced with respect to the BD ORTHO® orthophoto produced by the French National Geographic institute (IGN). The lighter part on the left corresponds to the drone-based orthophoto and the darker part on the right corresponds to the BD ORTHO®.
• Figure 4: Examples of useless tiles extracted from the aerial orthophoto of the Saint Leu lagoon in Reunion Island: (a) Example of a tile extracted from the aerial orthophoto of the Saint Leu lagoon in Reunion Island, with a high percentage of black pixels (b) Example of a group of tiles extracted from the aerial orthophoto of the Saint Leu lagoon in Reunion Island, with corresponding underwater images. The tiles in the middle do not have enough coverage of underwater images
• Figure 5: Footprint calculation of underwater images based on echosounder data, camera field of view and ASV angles.