Artificial Intelligence and Machine Learning in Water Resources and Environmental Engineering: A Systematic Review and Future Research Agenda

Abstract

Artificial Intelligence (AI) and Machine Learning (ML) are now leading-edge technologies being used to develop new ways to manage water resources and environmental engineering that outperform traditional statistical models by leveraging their capacity to manage nonlinear dynamics and large-scale multivariate datasets. The purpose of this systematic review is to determine the current status of AI/ML applications in the context of each method using a structured approach to synthesizing available evidence from primary databases such as Scopus and Web of Science. The methodologies used in this review are divided into a central taxonomy: predictive modeling using supervised learning; discovering patterns through unsupervised learning; making adaptive decisions using RL; and example hybrid models which use both analytical and data-driven algorithms.
There are many insights to be gained from this research study artificial intelligence(AI) has shown a high degree of accuracy in making flood predictions across various locations worldwide by means of Long Short-Term Memory neural networks (LSTM), with some reported values being nearly 99 % Nash-Sutcliffe Efficiency for specific locations in the tropics [1]; however, these findings must be interpreted carefully since they lacked adequate evaluation and were based only on one region where there was sufficient historical data, whereas this may not hold true in an area where little or no historical data exist and because these different types of locations of evaluating AI’s use. Along with AI, Internet of Things (IoT) devices have enabled real-time monitoring of water quality with close to .99 precision. AI is being used in the field of environmental engineering for automation of sorting waste with accuracy’s above 95% in the transition toward a circular economy. A significant hole remains in our knowledge regarding the advancement of AI/ML technology in water resources and environmental engineering. The vast majority of studies devoted to studying AI/ML in the water resources and environmental engineering industries utilize various AI/ML models to perform predictions of future conditions; however, few of these studies look at real-life applications for these AI/ML technologies or how they will be implemented at scale in countries that have limited access to quality data, such as Middle Eastern and dry developing countries. This gap was the motivation behind conducting this review; to consolidate evidence related to generalizability of model outputs, the need for open access data, and the movement from purely predictive models to decision-support systems. Nevertheless, critical challenges continue to exists such as reproducibility problems that arise from lack of freely available datasets/ code and lack of interpretability caused by complexity of deep neural net models. To resolve these drawbacks, the present review recommends: (i) To adopt PIML for improved accuracy through enforcing prior laws of physics within a ML model as well as increased clarity regarding the underlying mechanism of an ML prediction. (ii) Create a standardized data-sharing protocol to enable the use of benchmark datasets that are identified across the community. (iii) Use XAI techniques (SHAP and LIME) to better align an ML model’s predictions with howstakeholders are using those predictions for their decisions. (iv) Increase research on transfer learning strategies to improve the applicability of ML models in areas where there is limited data climate variation. All four of these approaches are the directions of future research proposed in this article.