We present results from extending that study to the regional scale spanning 338 locations across Sahel and East Africa in the Water Point Viewer (https://earlywarning.usgs.gov/fews/software-tools/25) of the United States Agency for International Development’s (USAID) Famine Early Warning Systems Network (FEWS NET), with some updates to the datastream and machine learning. We replaced the surface reflectance data stream from Landsat with that from Harmonized Landsat Sentinel-2 data to take advantage of the latter’s improved temporal sampling frequency. Additionally, to improve upon the weaknesses of the globular and constant-density cluster assumption, cluster instability, and noise points inclusion in the K-means clustering used in the earlier study, we considered more recent advanced clustering techniques like HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise).
We report on the improvement in clustering performance using both non-observation-based metrics (for example, SSE and silhouette score), and the validation performance metrics against selected observation points (for example, omission/commission errors and Normalized Mutual Information). Further, we discuss the trends in results due to variations in the waterpoint extents and the geography. The improved water extents product from this work form an input to a more realistic volumetric model by FEWS NET that replaces the existing depth based water model, towards better modeling and forecasting of water availability in pastoral regions.

