Mapping ecological indicators of human impact with statistical and machine learning methods: Tests on the California coast

摘要: Abstract Coastal ecosystems are exposed to multiple anthropogenic stressors such as fishing, pollution, and climate change. Ecosystem-based coastal management requires understanding where the combination of has large cumulative effects actions address impacts most urgently needed. However, on marine often non-linear interactive. This complexity is not captured by commonly used spatial models for mapping human impacts. Flexible statistical machine learning like random forests have thus been an alternative modeling approach identify important make predictions their combined effects. tests models' prediction skill limited. Therefore, we tested how well ten methods predicted three ecological indicators ecosystem condition (kelp biodiversity, fish biomass, rocky intertidal biodiversity) off California, USA. Spatial data representing ocean uses natural gradients were predictors. The errors estimated double block cross-validation. best achieved mean squared about 25% lower than a null model kelp biodiversity biomass; none worked biodiversity. general trends, but local variability indicators. For performing method was principal components regression. boosted regression trees. after tuning, this did include any interactions between stressors, ridge (a constrained linear model) performed almost well. While in theory flexible required represent complex stressor-ecosystem state relationships revealed experimental ecologists, with our data, flexibility could be harnessed because more overfitted due small sample sizes low signal-to-noise ratio. main challenge harnessing link obtaining suitable data. In particular, better describing temporal distribution We conclude discussing methodological implications future research.