A comprehensive review of AI-based methods used for forecasting ice jam floods occurrence, severity, timing, and location

River ice breakup can affect most rivers in cold climate during winter, posing a serious threat of Ice-Jam Floods (IJFs) to riverine communities. IJFs are challenging to predict due to their chaotic nature that arises from the complex interaction between hydroclimatic factors and river morphology. In addition, climate change has significantly impacted river ice patterns and the severity of IJFs in recent decades. However, recent advancements in computing power have led to the development of several Artificial Intelligence (AI) approaches to forecast IJF. Still, there is a lack of a systematic review that can adequately compare the different AI approaches together with the different hydrometeorological parameters used to forecast IJF. Therefore, the primary objective of this study is to review the various existing AI-based IJFs prediction models, their input parameters, and their potential strengths and limitations. The review showed that AI-based IJF prediction models can be grouped into four categories based on their objectives to forecast IJF occurrence, severity, timing, and location. The study also revealed that station-based data remained the primary source of information for predicting IJFs, but there has been a growing trend in recent years toward remote sensing, reanalysis products, and national databases, indicating their increasing prominence. Overall, air temperature, precipitation, and hydrometric parameters (discharge and water level) were the most frequently utilized input parameters. The review also categorized AI-based IJF forecasting models into four types: machine learning, hybrid, ensemble, and framework models. Although the framework approach has gained recent popularity in recent years, but still the machine learning and ensemble models were the most frequently used. While directly comparing the capabilities and limitations of different modeling approaches without considering the specific context of the sites in which they were applied can be misleading, several studies have demonstrated the potential of ensemble and hybrid approaches to improve model accuracy compared to single machine learning models. However, more studies are needed to confirm these conclusions.