Machine learning based mapping of physicochemical attributes in the Colombian Pacific seafloor

Abstract Seafloor mapping is a fundamental tool for sustainable use of hydrobiological resources, underwater research, and the management and conservation of oceans. Seafloor maps reveal a variety of biophysical features that enhance our understanding of ecological processes, facilitating monitoring, protection, and restoration actions. Seafloor mapping is expensive, hazardous, and complex, limiting its application especially in regions with scarce data. However, recent machine learning approaches are helping to overcome these challenges. Our objective was to map particle size, substrate hardness and organic matter content with a spatial resolution of 0.04° across the Colombian Pacific seafloor and down to a depth of 1000 m, by applying the XGboost algorithm. Feature engineering was applied to a set of 55 environmental predictors to train the algorithm and deploy it to the study area via regression-type predictions. Model performance measured as cross-validated-R² in test datasets was 0.86, 0.84, and 0.81 for particle size, substrate hardness, and organic matter content respectively. The results provide continuous information on the spatial configuration of seafloor attributes in the shallow waters of the Colombian Pacific, representing a significant advancement in the analysis and understanding of benthic ecosystems. The high coefficients of determination obtained support the accuracy of these methods, exceeding those of similar studies conducted in other regions. Uncertainty in predictions, both under and overestimations for the three variables are associated with spots with higher temporal variability in oceanographic conditions. Our maps provide a foundation for future research and the implementation of management and conservation policies for marine resources in the region.