GIScience & Remote Sensing (Oct 2019)
Using the random forest algorithm to integrate hydroacoustic data with satellite images to improve the mapping of shallow nearshore benthic features in a marine protected area in Jamaica
Abstract
Hydroacoustic and optical remote sensing have been commonly used to map shallow nearshore benthic features. However, the number, type, scale, and accuracy of the mapping products that can be obtained from the two sensors differ; as such, there can be limited agreement between their mapping products. These differences can be further accentuated if the hydroacoustic data are interpolated to produce a map. Interpolation introduces spatial uncertainty and reduces map accuracy. Consequently, maps generated from the two sensors may provide dissimilar spatial and temporal representations of the same benthic features. We therefore compared the performance of a random forest regression (RFr) and a universal kriging (UK) interpolation method and a post-classification enhancement that can be used to increase the accuracy and complementarity of benthic habitat maps derived from hydroacoustic data. First, we used single beam echosounder (SBES) survey bathymetry data from the Bluefields Bay marine protected area (MPA) in western Jamaica (13.82 km2 in size), to create a bathymetric surface model (BSM), from which rugosity and bathymetric position index (BPI) maps were generated. Next, the RFr was used to create submerged aquatic vegetation (SAV) percentage cover maps from the SBES SAV cover data by predicting cover at un-sampled locations. Predictors included auxiliary data such as depth, BPI, survey points coordinates and radiometrically corrected, deglinted and water column corrected image reflectance index values from each of the following: WorldView-2, Geoeye-1 and Landsat 8. Additionally, a SAV map was created using the UK. The most accurate SAV cover thresholds were identified and were used to create binary maps from the RFr and UK maps. A rugosity derived coral reef map was then added to the binary maps. The resulting benthic habitat maps had comparable accuracies and class coverage to benthic maps classified from GeoEye-1 and WorldView-2 images using pixel and object-based classifiers. However, map accuracies were calculated using a suboptimal number of reference points (<50) for two of the benthic map classes (SAV absent and coral reef). This was not considered to be problematic as the addition of the coral reef class to the binary maps resulted in a significant decrease in uncertainty (standard error and confidence interval width of the overall accuracy) and a significant increase in the user’s accuracy of the SAV absent map class. Also, the difference in uncertainty and accuracy between the map classes did not change. The methods used in this study can therefore be used to increase the accuracy (and to decrease the uncertainty) and the complementarity of maps derived from hydroacoustic data.
Keywords
