Ensemble Learning Enhanced Stepwise Cluster Analysis for River Ice Breakup Date Forecasting

Wei Sun, Qingjiang Shi, Yong Huang, Ying LvJournal of Environmental Informatics Letters 2019 · DOI

Abstract

Frequently occurring ice jams often cause concern in northern regions. Breakup timing is directly related to emergency responses preparation and thus its early accurate forecasting is beneficial to ice-related flooding management. The stepwise cluster analysis (SCA) is a non-parameter regression method, which generates a classification tree in the sense of probability through cutting or merging operations according to certain statistic criteria. To enhance SCA’s predictive performance, a SCA ensemble (SCAE) method is developed and applied to forecasting of annual river ice breakup dates (BDs). In detail, the SCA is employed as a base model at the lower level while the simple average method is selected as combining models at the upper level. The SCA base models are selected according to different performance selection criteria and searched for further combination. A site on a representative river prone to river ice flooding in Alberta, Canada is selected to demonstrate the effectiveness of the proposed SCAE. The results mainly show that: the SCA base models with multiple combinations of inputs and internal parameters are able to predict the BDs with good performances (the highest average of correlation coefficients for training can be 0.958); the optimal SCA base model has three inputs, which indicates that the temperatures before breakup and just after freeze-up as well as the maximum of water flow in March are relatively important indicators of BD. The optimal SCAE, including base models from different performance selection criteria, has the lowest average of root mean squared error, which improves upon the optimal SCA base model by 25.3%. It indicates the different model selection criteria do improve the diversity and thus further help to improve the performance of ensemble models. This first application of the SCAE to river ice forecasting highlights the possibility of using the ensemble learning paradigm to enhance the SCA. The potential applications of the SCAE to other forecasting problems are expected.

Introduction

In northern latitudes, the spring breakup of river ice is a high-stakes event. If ice breaks up too quickly or jams in narrow channels, it can trigger sudden, catastrophic flooding. Predicting exactly when this will happen is notoriously difficult because the process is highly nonlinear and influenced by a complex web of factors: air temperature, water flow, and upstream ice conditions. While traditional "black-box" models like Artificial Neural Networks (ANNs) can predict these dates, they often lack transparency and struggle with the small, "noisy" datasets typical of remote river stations. This research introduces a method that handles these complexities while remaining statistically robust and interpretable.

Methodology: Stepwise Cluster Analysis Ensemble (SCAE)

The core of this approach is Stepwise Cluster Analysis (SCA). Unlike standard clustering, SCA is a non-parametric regression technique that builds a classification tree. It works by iteratively "cutting" or "merging" clusters of data based on statistical significance tests (specifically Wilks’ likelihood-ratio and F-tests). This results in a tree structure where the "leaves" represent predicted breakup dates based on specific environmental conditions.

To further enhance predictive power, the researchers developed the SCA Ensemble (SCAE) framework:

  1. Base Learners: Dozens of SCA models are generated using different combinations of inputs and internal tuning parameters.
  2. Diversity Selection: Base models are selected using two different criteria—one focused on pure validation accuracy and another on the balance between training and validation performance—to ensure model diversity.
  3. Stacking: A Simple Average Method (SAM) combines the outputs of the selected base models into a final prediction.
Flowchart of the proposed SCAE forecasting framework.

Study Area and Data

The study focused on a high-risk site: the confluence of the Athabasca and Clearwater Rivers at Fort McMurray, Alberta, Canada. Historical breakup dates (BDs) spanning 1980 to 2015 were analyzed alongside 17 candidate indicators, including monthly maximum, minimum, and average temperatures from the previous fall through spring, and water flow rates from the Athabasca River. Due to the limited size of the dataset, Leave-One-Out Cross-Validation (LOOCV) was employed to ensure the model generalizes well to new years.

Key Results and Discussion

The optimal models consistently identified three primary drivers: temperatures just before breakup (March), temperatures immediately after freeze-up (the previous December), and the maximum water flow in March. This confirms that the "thermal memory" of the river and the mechanical force of spring flow are the dominant predictors.

The SCAE approach demonstrated a significant accuracy boost. The optimal ensemble model achieved a Root Mean Squared Error (RMSE) of 3.20 days, representing a 25.3% improvement over the best single SCA base model (RMSE of 4.28 days). Furthermore, the training phase reached correlation coefficients as high as 0.958, indicating the model captured the underlying physical relationships very effectively.

Comparison of observed vs. predicted breakup dates for a base SCA model (left) and the final SAM ensemble (right).

One of the most significant advantages of the SCAE is its "white-box" logic. Unlike opaque neural networks, the SCAE provides an explicit logic tree. This interpretability is vital for emergency managers who need to understand the environmental conditions driving a specific forecast.

Conclusion

The application of the SCAE to river ice forecasting highlights the potential of using ensemble learning paradigms to enhance traditional cluster analysis. By prioritizing model diversity over individual performance, the ensemble approach significantly reduced prediction variance and improved stability. Future work will explore more advanced ensemble techniques like boosting or bagging to further refine the SCA framework and apply it to other complex environmental forecasting challenges.

Summary updated

Abstract

Frequently occurring ice jams often cause concern in northern regions. Breakup timing is directly related to emergency responses preparation and thus its early accurate forecasting is beneficial to ice-related flooding management. The stepwise cluster analysis (SCA) is a non-parameter regression method, which generates a classification tree in the sense of probability through cutting or merging operations according to certain statistic criteria. To enhance SCA’s predictive performance, a SCA ensemble (SCAE) method is developed and applied to forecasting of annual river ice breakup dates (BDs). In detail, the SCA is employed as a base model at the lower level while the simple average method is selected as combining models at the upper level. The SCA base models are selected according to different performance selection criteria and searched for further combination. A site on a representative river prone to river ice flooding in Alberta, Canada is selected to demonstrate the effectiveness of the proposed SCAE. The results mainly show that: the SCA base models with multiple combinations of inputs and internal parameters are able to predict the BDs with good performances (the highest average of correlation coefficients for training can be 0.958); the optimal SCA base model has three inputs, which indicates that the temperatures before breakup and just after freeze-up as well as the maximum of water flow in March are relatively important indicators of BD. The optimal SCAE, including base models from different performance selection criteria, has the lowest average of root mean squared error, which improves upon the optimal SCA base model by 25.3%. It indicates the different model selection criteria do improve the diversity and thus further help to improve the performance of ensemble models. This first application of the SCAE to river ice forecasting highlights the possibility of using the ensemble learning paradigm to enhance the SCA. The potential applications of the SCAE to other forecasting problems are expected.
BibTeX
@article{sun2019ensemble,
  author = {Wei Sun and Qingjiang Shi and Yong Huang and Ying Lv},
  title = {{Ensemble Learning Enhanced Stepwise Cluster Analysis for River Ice Breakup Date Forecasting}},
  journal = {Journal of Environmental Informatics Letters},
  year = {2019},
  doi = {10.3808/jeil.201900005},
}
Figure