Seagrass Habitat Suitability Models using Multibeam Echosounder Data and Multiple Machine Learning Techniques

Abstract

Abstract Seagrass beds are important habitats in the marine environment by providing food and shelter to dugongs and sea turtles. Protection and conservation plans require detail spatial distribution of these habitats such as habitat suitability maps. In this study, machine learning techniques were tested by using Multibeam Echo Sounder System (MBES) and ground truth datasets to produce seagrass habitat suitability models at Redang Marine Park. Five bathymetric predictors and seven backscatter predictors from MBES data were used to representing topography features and sediment types in the study area. Three machine learning algorithms; Maximum Entropy (MaxEnt), Random Forests (RF), and Support Vector Machine (SVM) were tested. The results revealed that MaxEnt and RF models achieved the highest accuracy (93% and 91%, respectively) with SVM produced the lowest (67%). Depth was identified as the most significant predictor for all three models. The contributions of backscatter predictors were more central for SVM model. High accuracy models showed that suitable habitat for seagrass is distributed around shallow water areas (<20 m) and between fringing reef habitats. The findings highlight that acoustic data and machine learning are capable to predict how seagrass beds are spatially distributed which provide important information for managing marine resources.