POSTER SESSION 2
10:30–12:30, Tuesday, September 15
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
ABSTRACT 1407 | POSTER 10
CORAL BLEACHING DETECTION FROM PLANETSCOPE IMAGERY USING A CONVOLUTIONAL NEURAL NETWORK CLASSIFIER
Coral reefs support approximately 25% of marine species and sustain the livelihoods of hundreds of millions of people globally, yet large-scale reef health monitoring remains constrained by costly, spatially sparse field surveys. Existing satellite-based products detect thermal stress rather than bleaching directly. The 2023–2025 global bleaching event, the most extensive on record, affected approximately 84% of the world’s coral reef area, underscoring the urgent need for scalable optical detection methods. Here, we trained a pairwise ResNet-50 convolutional neural network (CNN) on PlanetScope SuperDove imagery (3 m resolution) to classify 120×120 pixel coral patches as healthy or bleached across 12 reef sites spanning four ocean basins. Bleached corals exhibit elevated reflectance in the blue (492 nm) and green (566 nm) bands, and we asked whether a CNN classifier could learn these spectral and spatial signatures to generalize across reef regions. Using a leave-one-site-out validation design, the model achieved an average accuracy of ~90% on reef sites withheld entirely from training. Spectral importance analysis confirmed that these same spectral bands were the strongest predictors of bleaching status, and Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations verified that model attention concentrated on coral structures rather than surrounding substrate. As marine heatwaves increase in frequency and severity, these results demonstrate that multispectral CNNs offer a promising, scalable pathway for reef health monitoring; additional training data across diverse regions will be needed to confirm global transferability.
Mariam Ayad*, University of California, Santa Cruz, [email protected], https://orcid.org/0000-0003-3590-9348
Kevin Valencia, University of California, Los Angeles, [email protected]
Christine Lee, Jet Propulsion Laboratory, [email protected], https://orcid.org/0000-0003-1615-236X
Hannah Druckenmiller, California Institute of Technology, [email protected]
Raphael Kudela, University of California, Santa Cruz, [email protected], https://orcid.org/0000-0002-8640-1205
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
