1
HYDROINFORMATICS AND APPLICATIONS OF ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING IN WATER-RELATED PROBLEMS

Gerald A. Corzo Perez1 and Dimitri P. Solomatine1,2,3

1IHE Delft Institute for Water Education, Delft, The Netherlands

2Water Resources Section, Delft University of Technology, Delft, The Netherlands

3Water Problems Institute of the Russian Academy of Sciences, Moscow, Russia

In recent years, there has been a surge of interest in machine learning (ML) and artificial intelligence (AI) due to the effectiveness of deep learning algorithms and the increasing availability of large data sets. This chapter provides a brief overview of the applications of AI and ML techniques in hydroinformatics, a field that deals with advanced information technology, data analytics, and modeling for aquatic environment management. Data-driven models are becoming more common in water management as they can reveal hidden patterns in data and offer improved accuracy in certain situations. This chapter highlights the importance of spatiotemporal data analysis, pattern recognition, and optimization approaches in water resources management under uncertainty. It does not offer a comprehensive review of all methods but rather focuses on selected ML techniques widely used in water-related problems. Additionally, the chapter discusses the challenges associated with using ML models, such as black-box criticisms, and the potential of hybrid models that combine the strengths of ML and physically based process models for more robust solutions in hydroinformatics.

1.1. Introduction

Hydroinformatics deals with advanced information technology, data analytics, modeling, artificial intelligence (AI), and optimization applied to problems of aquatic environment for the purpose of informing management. Many of these technologies have become standard tools that support water management decisions around the world. However, the technologies are developing further, new ones are emerging, and this allows for applying them to more complex and interesting problems. One can find multiple examples when environmental and hydrological problems have been dealt with not only by employing physically based (process) models, but also advanced data analysis tools and machine learning models have been used. Using AI techniques in geosciences has a long history. Hydroinformatics, formulated by Abbott (1991) 30 yr ago, has been defined as a union of computational hydraulics (CH) and AI (so that HI?=?CH???AI), and during the last three decades we have been witnessing a much wider use of AI, with a large number of successful practical applications. The first stage of such development has been covered, for example, in the edited volume Practical hydroinformatics: Computational intelligence and technological developments in water applications (Abrahart et al., 2008), and in dozens of other books and hundreds of research papers covering these new developments.

Currently, we see a new wave of interest in machine learning (ML) and AI, which is partly explained by the demonstrable effectiveness of the new generation of deep learning algorithms and availability of large data sets (see, e.g., Nearing et al., 2021), and this brings new possibilities for hydroinformatics research and practice. With an increasing amount of data collected about the environment, physically based models are more and more complemented and sometimes even replaced by data-driven models. Lacking the ability of physically based models to explain the physics of underlying processes, data-driven models are however able to discover the hidden patterns in data and often can be more accurate, and play an important supporting role, in water management. Pattern recognition (e.g., automatic identification of flooded areas on satellite images) has been one of the main tasks solved by machine learning, and lately has been given an additional push by the development and use of deep learning, an important class of machine learning algorithms, and of AI in general. Data analytics plays an important role in water resources when data are multidimensional, and spatial and time dimensions have to be dealt with in a coordinated fashion. In relation to water resources, both dimensions were always important, but recently the need to handle huge amounts of remote sensing data ("big data") has become more pronounced. These developments have motivated new research efforts in the context of predicting hydrological extremes and call for testing novel approaches of spatiotemporal data analysis and machine learning. Due to much easier access to supercomputing facilities, there are increased possibilities to study the models uncertainty (typically using Monte Carlo frameworks), and machine learning can also play a role in building predictive models of such uncertainties. An issue in water resources management is optimal planning and operation under uncertainties, and this is where the role of AI-driven approaches is also becoming more important. Classical optimization approaches (gradient-based nonlinear optimization) typically cannot help much, since such optimization is model based, and objective functions (and their gradients) cannot be analytically expressed. Optimization approaches developed under the framework of computational intelligence (various types of randomized search, e.g. evolutionary approaches) have been the focus of hydroinformatics for three decades, but the new problems and the increased data availability lead to the necessity of testing new approaches and their critical analysis.

This chapter aims at presenting a brief overview of AI- and ML-related building processes and methods widely used for water-related problems, in the context of the chapters presented in this volume. AI is a concept that covers a wide area of science and technology, however, quite often it is used interchangeably with ML, which is in fact a narrower notion. One may find in literature quite a large number of AI- and ML-related subareas: big data, data mining, pattern recognition (PR), natural language processing (NLP), neural networks, deep learning, and so on. We will not go into a discussion about terminology and differences in AI and ML; for the purpose of this chapter and the issues covered in the book, it would be right to use a somewhat narrower term, that is, machine learning.

ML techniques have been widely used in water resources during the last decades, however, at the same time, one may observe also inadequate use of ML-related modeling procedures, unjustified selection of algorithms, and even lack of understanding of why a model provides good or poor performance in mathematical and statistical sense. There is also well-known criticism of ML and statistical techniques by practitioners who are used to employing physically based (process) models; they are pointing out that a water resources problem interpretation is hidden in the so-called black box of a ML model. There is indeed a challenge of posing the problem in the right way: how domain knowledge can drive selection, building, and tuning a ML model. Lack of data and its uncertainty also makes it difficult for practitioners to feel confident about ML models.

On the other hand, the strength of ML is in its ability to represent the relationships between inputs and outputs, provided enough data are available. Although the relatively recent advances in deep learning have opened the door to the new ways of using spatiotemporal data, and at the same time motivating new algorithm developments from spatial patterns and, in general, all types of computer vision algorithms, not all problems can be tackled by ML. Input and output relations can be so complex that ML techniques may not be able to find the hidden patterns, and in such cases hybrid models, combining power of ML and process models (so-called physics-aware AI; see, e.g., Jiang et al., 2020) would be needed. Such hybrid approaches are given now increased attention in hydroinformatics.

This chapter is not intended to provide a comprehensive review of methods (which are covered in hundreds of books and in the referred literature herein), but rather focuses on some important elements of ML model building, and presents basics of several selected ML techniques quite widely used in solving water-related problems, allowing for "feeling the flavor" of ML.

1.2. Key Principles of ML/Hydroinformatics

1.2.1. AI and ML Definitions

There is a large number of evolving definitions of AI, and this can be explained by its permanent evolution and shifts in priorities, and the advances in the used mathematical instruments. Many literature sources point out that for the first time the term AI was used in 1956 at the Dartmouth Conference, were John McCarthy, Alan Turing, and other founding fathers of AI, help to coin the term artificial intelligence. One of the definitions reads: "AI is the field devoted to building artificial animals (or at least artificial creatures that, in suitable contexts, appear to be animals) and, for many, artificial persons (or at least artificial creatures that, in suitable contexts, appear to be persons)" (Stanford Encyclopedia of Philosophy, 2018). On the other hand, Wikipedia defines it as the "intelligence demonstrated by machines, unlike the natural intelligence displayed by...