Toward Bridging the Gap Between Data-Driven and Mechanistic Models: Cluster-Based Neural Networks for Hydrologic Processes

Abstract

The emergence of artificial neural network (ANN) applications in hydrology, among other data-driven techniques, has created a new chapter in this field of science that has been termed “neurohydrology” (Abrahart, 1999). However, the number of operational solutions is still limited because most of the practitioners and many researchers have difficulty accepting ANNs as a standard technique. This chapter presents an attempt to bridge the gap between ANN models and mechanistic hydrologic models using cluster-based ANN models. A novel spiking modular neural network (SMNN) model is proposed, configured, and trained in a way that assigns different sub-components of the hydrologic processes to corresponding sub-components of the developed SMNN model. The modular nature of the SMNN helps to find domain-dependent relationships. The proposed model configuration reflects the ability of the ANN technique to mimic mechanistic models in terms of their sensible internal structure and the way they model various hydrologic processes. The meteorological data, including air temperature, ground temperature, wind speed, relative humidity, and net radiation, from a semi-arid region in northern Alberta, Canada, are used in this chapter to estimate actual evapotranspiration (AET). The estimated AET is contrasted against the values measured by an eddy covariance system. The results of the SMNN model, which uses an unsupervised clustering technique to group the input data into a definite number of clusters, are compared with the results of another modular ANN model that relies on a supervised clustering technique, and with a traditional global feedforward ANN model. The estimates of AET using the Penman–Monteith model are also presented for reference and comparison purposes. The SMNN model provided the best performance and was shown to be effective in discretizing the complex mapping space into simpler domains that are easier to learn.