Mapping emergent coral reefs: a comparison of pixel‐and object‐based methods

Although emergent coral reefs represent a significant proportion of overall reef habitat, they are often excluded from monitoring projects due to their shallow and exposed setting that makes them challenging to access. Using drones to survey emergent reefs overcomes issues around access to this habitat type; however, methods for deriving robust monitoring metrics, such as coral cover, are not well developed for drone imagery. To address this knowledge gap, we compare the effectiveness of two remote sensing methods in quantifying broad substrate groups, such as coral cover, on a lagoon bommie, namely a pixel-based (PB) model versus an object-based (OB) model. For the OB model, two segmentation methods were considered: an optimized mean shift segmentation and the fully automated Segment Anything Model (SAM). Mean shift segmentation was assessed as the preferred method and applied in the final OB model (SAM exhibited poor identification of coral patches on the bommie). While good cross-validation accuracies were achieved for both models, the PB had generally higher overall accuracy (mean accuracy PB = 75%, OB = 70%) and kappa (mean kappa PB = 0.69, OB = 0.63), making it the preferred method for monitoring coral cover. Both models were limited by the low contrast between Coral features and the bommie substrate in the drone imagery, causing indistinct segment boundaries in the OB model that increased misclassification. For both models, the inclusion of a drone-derived digital surface model and multiscale derivatives was critical to predicting coral habitat. Our success in creating emergent reef habitat models with high accuracy demonstrates the niche role drones could play in monitoring these habitat types, which are particularly vulnerable to rising sea surface and air temperatures, as well as sea level rise which is predicted to outpace reef vertical accretion rates.