Traditional Methods

Rivers have been an important source of sustenance and means of transportation for humans for thousands of years, stretching back at least to ancient Egypt and other riverine civilizations of the same era. The Palermo Stone, which at over 4,000 years old is one of the oldest surviving records from ancient Egypt, lists annual measurements of Nile river flooding, which was critical to the health of the river and therefore the Egyptian civilization.¹

Understandably, forecasts of floods and dry spells have been greatly sought-after for thousands of years. While the earliest forecasts likely would have been derived from the secret divinations of priests,² today we have more reliable and sophisticated methods at our disposal. Traditional methods of river level forecasting rely on the construction of a physical model of a hydrological basin, followed by the use of that model to translate a weather forecast into a forecast for the river level. These models, however, can require significant computational resources, and nonlinear responses to environmental conditions can be difficult to model successfully.³

In recent years, machine learning methods have gained significant traction in this field. They are capable of learning non-linear systems empirically, typically require significantly fewer computational resources once trained, and are not bound to a predefined input-response model.