Sammendrag
The rapidly expanding field of machine learning (ML) provides many methodological opportunities which match very well with the needs and challenges of hydrological research. Due to extended measurement networks, more frequent automatic measurements of hydrological variables, and not the least increasing use of remote sensing products, the era of big data surely has arrived in hydrology. Process-based models are usually developed for certain spatiotemporal scales, not fitting easily to the scope of the new datasets. Automatic methods that learn patterns and generalizations have been demonstrated to be superior in many applications. The chapter provides an overview of some of the most important machine learning algorithms which have been used in the hydrological literature. It will be shown that there is no single best method among them, but instead a spectrum of methods should be utilized, from highly flexible ones to more parsimonious learning methods, depending on the specific hydrological application, research question, and data availability. Most machine learning techniques require a calibration and a validation dataset for training. As these data are usually correlated in time and space, the problem of bias-variance tradeoff arises will be discussed as a simple example. The presentation of ML algorithms, roughly following chronological order, is discussed starting with artificial neural networks through support vector machines to gradient boosting machines. As data streams increase, these and other machine learning techniques will play an ever more important role in hydrology.
Vis fullstendig beskrivelse
