Convolutional neural network and long short-term memory models for ice-jam predictions.

• Published in: The Cryosphere
• Main Authors: Madaeni, Fatemehalsadat, Chokmani, Karem, Lhissou, Rachid, Homayouni, Saeid, Gauthier, Yves, Tolszczuk-Leclerc, Simon
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: 2022
• Subjects: artificial neural network; flooding; hydrometeorology; numerical model; prediction; Canada; The Cryosphere
• Online Access: https://espace.inrs.ca/id/eprint/12728/; https://doi.org/10.5194/tc-16-1447-2022

In cold regions, ice jams frequently result in severe flooding due to a rapid rise in water levels upstream of the jam. Sudden floods resulting from ice jams threaten human safety and cause damage to properties and infrastructure. Hence, ice-jam prediction tools can give an early warning to increase response time and minimize the possible damages. However, ice-jam prediction has always been a challenge as there is no analytical method available for this purpose. Nonetheless, ice jams form when some hydro-meteorological conditions happen, a few hours to a few days before the event. Ice-jam prediction can be addressed as a binary multivariate time-series classification. Deep learning techniques have been widely used for time-series classification in many fields such as finance, engineering, weather forecasting, and medicine. In this research, we successfully applied convolutional neural networks (CNN), long short-term memory (LSTM), and combined convolutional-long short-term memory (CNN-LSTM) networks to predict the formation of ice jams in 150 rivers in the province of Quebec (Canada). We also employed machine learning methods including support vector machine (SVM), k-nearest neighbors classifier (KNN), decision tree, and multilayer perceptron (MLP) for this purpose. The hydro-meteorological variables (e.g., temperature, precipitation, and snow depth) along with the corresponding jam or no-jam events are used as model inputs. Ten percent of the data were excluded from the model and set aside for testing, and 100 reshuffling and splitting iterations were applied to 80g% of the remaining data for training and 20g% for validation. The developed deep learning models achieved improvements in performance in comparison to the developed machine learning models. The results show that the CNN-LSTM model yields the best results in the validation and testing with F1 scores of 0.82 and 0.92, respectively. This demonstrates that CNN and LSTM models are complementary, and a combination of both further improves classification.