Coral reef detection using ICESat-2 and machine learning

Abstract

As anthropogenic impacts threaten natural habitats, effective monitoring strategies are crucial. Coral reefs, among the most vulnerable ecosystems, traditionally employ monitoring techniques that are labor-intensive and costly, prompting the exploration of remote sensing as a cost-effective alternative. Launched in October 2018, the Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) provides high-resolution, high-frequency data, with its green laser offering unprecedented opportunities for bathymetric and coral reef applications. This study investigates the use of ICESat-2 data for atoll coral reef detection, utilizing Heron Island in the Great Barrier Reef, AU, and employing machine learning models. A binary logistic regression (BLR) model and convolutional neural network (CNN) were tested for determining coral reef presence, with the CNN outperforming the BLR in accuracy (85.4%), F1 score (43%), and false positive rate (13.1%). A challenge of the study included the difficulty of balancing false positive rates in predictive models to avoid over- or underestimations of reef extent. These obstacles were mitigated through the integration of algorithmically derived pseudo-rugosity and slope metrics as innovative proxies for seafloor complexity, significantly improving predictive performance. Feature importance analysis identified satellite-derived bathymetry (SDB) depth as the most critical predictor of coral presence, followed by pseudo-rugosity, slope, and various other depth measurements. This research establishes a new application of ICESat-2 data combined with advanced machine learning techniques as a promising method for efficient and cost-effective coral reef monitoring. Future work should refine algorithms and incorporate additional environmental variables to improve model performance across various reef types.