Forecasting river sediment deposition through satellite image driven unsupervised machine learning techniques

Abstract

The hydrologic and riverine systems of the Earth's surface, subject to ever increasing instability, requires forecasting techniques now more than ever. Existing mathematical and machine learning models for forecasting river sediment deposition are driven by non-remotely sensed data and/or require ground truth data. The study aims to demonstrate that remote sensing and unsupervised machine learning techniques coupled with appropriate validation metric can be employed to quickly forecast regions that are subject to future river sediment deposition. The results indicate that the NIR and other bands outside the visible range are determinant features for forecasting depositional areas, a multi-algorithm approach is both successful and necessary in indicating such regions and appropriate learning algorithm selection is more dependent on image cluster interpretability rather than cluster validation indices. However, the Silhouette Width validation index consistently picks the algorithm providing the result with most interpretability. Regions to be subjected to river sediment deposition were confirmed when validated against satellite images of the same region captured in the future. Such techniques can be employed in policymaking, disaster management, labeling for supervised learning etc.