A Deep Learning Approach for Modeling and Hindcasting Lake Michigan Ice Cover

Abstract

In large lakes, ice cover plays an important role in shipping and navigation, coastal erosion, regional weather and climate, and aquatic ecosystem function. In this study, a novel deep learning model for ice cover concentration prediction in Lake Michigan is introduced. The model uses hindcasted meteorological variables, water depth, and shoreline proximity as inputs, and NOAA ice charts for training, validation, and testing. The proposed framework leverages Convolution Long Short-Term Memory (ConvLSTM) and Convolution Neural Network (CNN) to capture both spatial and temporal dependencies between model input and output to simulate daily ice cover at 0.1{\deg} resolution. The model performance was assessed through lake-wide average metrics and local error metrics, with detailed evaluations conducted at six distinct locations in Lake Michigan. The results demonstrated a high degree of agreement between the model's predictions and ice charts, with an average RMSE of 0.029 for the daily lake-wide average ice concentration. Local daily prediction errors were greater, with an average RMSE of 0.102. Lake-wide and local errors for weekly and monthly averaged ice concentrations were reduced by almost 50% from daily values. The accuracy of the proposed model surpasses currently available physics-based models in the lake-wide ice concentration prediction, offering a promising avenue for enhancing ice prediction and hindcasting in large lakes.