Comparative analysis of Random Forest and Support Vector Machine for benthic habitat segmentation

Escalating climate impacts prompt governments to act as seen in the fifth Conference of the Parties (COP), demanding eco-friendly practices to limit warming to 1.5 degrees C. Carbon accounting is vital for global sustainability, requiring robust national monitoring of stocks and emissions. Remote sensing technology and satellite data enable modeling terrestrial carbon reserves, though challenges remain for coastal areas due to water attenuation. Ongoing studies aim to prove the technology's viability, despite accuracy issues in capturing shallow coastal environments. With this being gap, this study developed a methodology to map a coastal environment using satellite data and machine learning. Sentinel-2 MSI, an open-source multispectral image, was utilized in this study. Geospatial derivatives such as ratios of the visible bands, bathymetry model using the Stumpf's ratio and principal components which contained at least 90% of uncorrelated data were also integrated in the modeling process to improve benthic feature separability. Different combinations of the datasets were also explored in this study. Benthic habitat models were produced using Random Forest (RF) and Support Vector Machine (SVM) machine learning algorithms for each variable combination. The generated models generated overall accuracies ranging from 0.69 to 0.74 and 0.22 to 0.68 respectively. This translated to a maximum percent difference of 77% for the case of RGB model only and a minimum of 8% using all the variables. In terms of using different variable combinations, RF exhibited robust performance showing relatively consistent results compared to SVM which produced a wide range of accuracy values across the different models.