Uncertainty analysis and regression models for marine human impact mapping
People use and affect the oceans in many ways. This has resulted in a world-wide degradation of coastal and marine ecosystems. The need for better spatial management of threats like fishing and pollution has led researchers to develop maps of human impact on marine ecosystems. Yet these maps are based on simple spatial models, whereas experimental research shows that the effects of multiple stressors on marine organisms, populations and ecosystems are complex. This dissertation thus seeks to improve the methodological and conceptual foundations of spatial human impact modeling by advancing two parallel lines of research. First, I use computational experiments to investigate uncertainty related to model assumptions and data quality in regional and global maps created with a widely used spatial human impact model. This research shows that sources of uncertainty which alone have only small effects on human impact maps, have large effects in combination. However, I also identify spatial patterns of human impact that are robust, i.e. that are consistently found under a wide range of model assumptions and that are insensitive to data quality. Second, as an empirical alternative to established models, I use machine learning methods to train and test regression models that predict indicators of marine ecosystem condition based on human stressor maps, satellite images and other spatial data. This research shows that machine learning methods can generate accurate maps of ecological indicators, but also that the availability of suitable data is a major barrier to using machine learning for human impact mapping to its full potential.