Lake Ice Phenology Extraction using Machine Learning Methodology

Abstract

Lake-ice phenology is one of the key cryosphere indicators. Based on changes in time series of daily lake ice coverage and the microwave brightness temperature, two machine-learning methods were selected for estimating the phenology of lake ice. Using MODIS lake ice coverage time series, a convolutional neural network (CNN) was used to automatically classify lakes as either having or not having regular freezing signatures. The classification results were compared to the test samples and classification accuracies of 91% and 100% were found for lakes that do and do not freeze annually, respectively. In order to extract the lake ice phenology from passive microwave brightness temperature data, support vector regression (SVR) was used: the freeze-up start (FUS) and break-up end (BUE) were extracted using time series of estimated emissivity. The lake ice phenology results obtained using SVR were compared with the reference results. The values of R2 for FUS and BUE were 0.8928 and 0.8899, respectively. These results show that it is possible to use AI methods for the fast extraction of the phenology of lake ice and that these methods can be applied to lakes globally.