Hydroacoustic sensors are better suited for mapping shallow turbid coastal waters, but surveys must take place along transects, necessitating the interpolation of processed, large, spatially dependent hydroacoustic datasets to produce continuous maps. The aim of this study was to evaluate the performance of interpolation techniques that can handle data obtained from a hydroacoustic survey (covering 23 km2) conducted in 2016 in a shallow bay in southwestern Jamaica. Independent data or a 5-fold cross validation were used to compare the accuracies of continuous maps of benthic sediment percentage composition and submerged aquatic vegetation cover (SAV), interpolated from the survey data using random forest (RF) regression (RFr), RFr kriging (RFrk) and Ensemble Machine Learning (EML) with rotated/oblique coordinates. Categorical/thematic maps of sediment types interpolated using RF (RFc) and EML classifiers and a spatial Markov chain model with indicator kriging predictor/simulation (Mc-IK) were also compared. RFr, RFc, RFrk and EML covariates included coordinates and auxiliary data. The following maps were more accurate than corresponding interpolated maps: EML - mud, sand, silt, clay and SAV; RFr - loss on ignition [LOI]; RFrk - calcium carbonate; Mc-IK – sediment types with 3 and 4 classes. The Mc-IK maps were also the most suitable maps for assessing spatial and quantitative changes. The predominant substrate type in the bay was muddy sand covering 15.7 km2. The highest percentage composition of mud, silt, clay, and organic matter [LOI] was found close to the mouth of the Black River where sandy mud was found, indicating a high input of terrigenous sediment. Calcium carbonate composition showed an opposite spatial trend and was positively associated with higher SAV percentage cover. The bay is now part of a protected area, as such, these maps will be used as a baseline to gauge the impact of increased protection/management on benthic dynamics.
Read full abstract