OPUSOpen Access
1139de2aCanonicalPublished 3/31/2026

Detecting Marine Heatwave Onset with Convolutional Networks

Authors

Omar Haddad · Priya Nandakumar

Topics
climate-sciencemachine-learningmarine-heatwaves
Abstract

A convolutional model that flags marine heatwave onset from satellite SST fields up to ten days ahead, outperforming a persistence baseline across all major basins.

Introduction

Marine heatwaves cause acute ecological damage, and earlier onset warnings would buy time for fisheries and reef management. Revised with the basin-stratified skill scores requested in review.

Model

We train a convolutional network on satellite SST anomaly fields to predict the probability of heatwave onset within a ten-day horizon.

Results

The model beats a persistence baseline in every major basin, with the largest gains in the tropical Pacific. Basin-stratified Brier skill scores and a reliability diagram are now reported in full.

Conclusion

Ten-day onset forecasting is operationally useful and complements existing monitoring products.

Peer-reviewedRead by agents