Comparative analysis of contemporary remote sensing techniques and species distribution modelling in predicting spatial distribution of coastal habitats

Statistical and machine learning models, trained with in situ species observations and continuous covariate data, are commonly used to produce marine habitat maps by predicting ecological features over large areas. Here we systematically compare such approaches based on species distribution models (SDM) and more recent remote sensing (RS) models for mapping coastal habitats spanning seven European marine regions. We implemented each model using conventional approaches and data on species occurrences as well as covariates from 1) environmental maps (SDM), 2) Sentinel-2 satellite images and bathymetry (RS), and 3) commercial very high resolution multispectral images and bathymetry (RS – only two areas). We measured each model's predictive performance using cross-validation and analysed how it changed with data quality, depth and water opacity. The SDM approach generally outperformed the Sentinel-2-based RS approaches, being superior in five out of seven areas. However, SDMs performed poorly in the Mediterranean Sea likely due to small variability in the available environmental covariates. RS showed promise for producing benthic habitat maps even though it required larger in-situ datasets and reusing historical survey data proved more challenging than with SDMs. The performance of the SDMs depended on the availability of high-quality, ecologically relevant covariate maps and the taxonomic resolution of the in-situ data. The present study offers an in-depth comparison of the advantages and disadvantages between the approaches, especially on how they perform across different areas for coastal habitat mapping, providing practical guidance on their use, as well as suggestions for further development and co-use.