Real-time prediction of river ice breakup phenomena: A jittered genetic programming model and wavelet analysis integrating remotely sensed imagery and machine learning

Abstract

Forecasting the timing of breakup ice jams in rivers is crucial for early flood warning and effective management in cold regions where rising river flows can lead to significant damage. However, this task is hindered by insufficient data and the complex dynamics of river ice. These obstacles pose challenges in developing precise forecasting models. This study aimed to address the data scarcity issue by introducing innovative machine learning methods, focusing on classification and jittering (J), binary genetic programming (BGP), and wavelet transform (WT) for river ice forecasting (WT-JBGP) in the Tornio River, situated between Finland and Sweden. By considering time scales ranging from 2 to 32 days and time lags of 1 to 3 days, this method was applied to enhance the predictive capabilities of predictors. The findings reveal that certain predictors, with specific time scales and time lags, significantly influence the timing of breakup events. These include the 8-day temperature and 32-day discharge, both with a 2-day lag time, as well as the 4-day precipitation, approximation of albedo, and 16-day atmospheric pressure at ground level, all with a 1-day lag obtained from ERA5 and recorded data. Additionally, we conducted a quantitative evaluation of the effectiveness of the proposed model and contrasted its efficacy with that of BGP, WT-BGP, and advanced J-BGP techniques. The WT-JBGP model attains the highest overall classification accuracy of approximately 0.91, alongside a Heidke Skill Score exceeding 0.78, and a Positive Predictive Value surpassing 0.85, thereby demonstrating its superiority over competing methodologies. In summation, this study offers a promising approach to overcoming observed data scarcity in breakup date prediction, providing valuable insights into river ice dynamics.