Long-term monitoring and forecasting of glacial lake dynamics using Landsat time series data, Google Earth Engine, machine learning, and geospatial analysis

Long-term simultaneous tracking of the dynamics of regional glacial lakes can help in identifying lakes prone to climate change and Glacial Lake Outburst Flood (GLOF). This study demonstrates a methodology of tracking the dynamics of glacial lakes using Landsat time series data and SRTM DEM of Sikkim Himalaya over three decades (1987–2020), using a random forest classifier (RFC) and an artificial neural network (ANN). The classifiers were trained with features like slope, hillshade, automated water extraction index, band ratio, modified normalized difference water index, normalized difference water index, and water ratio index. The performance of the classifiers were measured using parameters like Accuracy, Kappa, Sensitivity, Specificity, Precision, F1 score, and Area under the curve. Furthermore, imbalance tests were performed to validate the predictions of the classifiers. On average, RFC marginally outperformed ANN with an accuracy of 98%. The slope was the most important determinant in mapping the glacial lakes, followed by the automated water extraction index. Time series data generated from this method was used in forecasting the fate of numerous glacial lakes of the Sikkim Himalaya. Models like Brown’s and Holt’s exponential smoothing, and the random walk model were applied for forecasting. The forecasts were validated with varying degrees of accuracy. The proposed methodology helps in overcoming the challenges of mapping glacial lakes over a vast geographic area over a prolonged period and generates time series data of their spatial extent. The methodology demonstrated here will be useful for the long-term mapping and monitoring of glacial lakes all over the world.