Parametric versus nonparametric machine learning modelling for conditional density estimation of natural events: Application to harmful algal blooms

Besides the complex effect of global warming on extreme events, spatiotemporal variability of natural phenomena often carries the legacy of anthropogenic activities. Moreover, any feedback induced by these activities on climate brings additional complexity when modelling natural events. For extreme values, climate or physiographic patterns often induce non stationarity, or long-term changes. In this context, parametric models may become inadequate given the complexity of the studied phenomena and their systematic changes through space and time. In this paper, we assess the use and ensuing efficiency of nonparametric machine learning (npML) methods to estimate and predict extreme values associated with natural events. These npML methods are compared to a commonly used parametric machine learning (pML) approach, the nonstationary frequency analysis model. We use a historical database compiling the frequency of harmful algal blooms (HAB) in Que & PRIME;bec, Canada. Results show that a 19-covariate RFCDE model leads to the best mean estimate among the considered models. However, for low and large quantiles, the 4-covariate RCDE model provides better agreement between observed and simulated bloom frequencies. The models may be used to assess the effects of climate change and anthropogenic developments on the frequency of HAB. They may also be leveraged to measure the efficiency of mitigation scenarios and to identify priority areas for restoration plan strategies. Recommendations are finally made regarding the estimation of the conditional density to predict extreme values associated with natural events.