A remote sensing model for coral recruitment habitat

Abstract

1. There is an increased need for habitat models of coral reefs to aid active interventions and restoration efforts rather than just map reef habitat extent. Habitat models provide the base layer for connectivity estimates among coral populations, to identify sites with likely increased coral survival and reduced project costs. We present a spatial modelling method to better identify coral reef habitats for coral recruitment and survival. 2. We demonstrate these methods for an offshore reef cluster in North Western Australia that has lost live coral as a result of successive bleaching events. This isolated set of reefs has high levels of coral self-recruitment, which makes them a good test for methods that underpin spatial optimisation of coral restoration locations. 3. Our method uses satellite derived, multiscale depth and rugosity metrics to model habitat classes, and applies randomForest cross-validated variable importance to optimally select a subset of these metrics that best model each habitat. Our approach produces improved models compared to existing object-based segmentation methods by overcoming a bias toward larger single-scale homogenous reef features. 4. Our method produces robust models of target reef recruitment habitats with encrusting and juvenile corals with crustose coralline algae but also models of habitats that are not desirable as recruitment habitat such as turf algae, sand, and rubble. All these models meet or exceed accepted accuracy benchmarks (accuracy and AUC values >0.8 on blind hold-out data used for validation and kappa values >0.6). 5. These models are specifically designed to support coral restoration including prime coral recruitment habitats with a mixture of crustose coralline encrusting and juvenile corals. 6. We demonstrate how spatially explicit model validation can identify locations within the target area where models are most and least supported, providing valuable additional insights into planning targeted survey efforts to reduce spatial uncertainty, and targeted restoration interventions that have the highest probability of success. 7. Our method has broad utility due to the widely and freely available Sentinel 2 remote sensing data cover for most reef systems globally, and field data collected with a simple drop camera system and classified with via the ReefCloud open access image classification system. It also has potential for adaption and refinement using other finer scale commercial remote sensing satellite data that can be used to drive digital elevation models including Planet Labs Skysat and Worldview 3 data.