Leveraging satellite data and machine learning to quantify rooftop rainwater harvesting reliability and tank requirements for schools and health care facilities in Uganda

Benjamin Okello , Patrick Oler

Many schools and health facilities in Uganda lack a reliable water supply, yet most have roofs that could collect rain. For 41,592 facilities (32,881 schools and 8,711 health facilities), we estimated how much water their roofs can supply, how large a tank each one needs, and how often that supply fails. Roof areas came from AI-derived building footprints (Google Open Buildings v3) and daily rainfall for 1981 to 2025 from CHIRPS. A daily yield-after-spillage tank model gave the supply available on 90% of days and the smallest tank that meets each facility's demand on 90% of days. On a standard 200 m² roof, a 10 m³ tank supplied a median of 359 litres per day on 90% of days, from 255 L in the Northern region to 411 L in the Eastern region; at the same annual rainfall, Northern facilities received about 40% less because their rain falls in one long season. With mapped roofs, 77.1% of schools and 62.2% of health facilities could reach the target with a tank of 200 m³ or less, and among these the median school needed 39.5 m³ and the median health facility 5.0 m³. The school share ranged from 42.3% to 92.8% across enrolment scenarios and fell to 44.1% if only buildings within 60 m are connected and to 37.8% if only 30% of the mapped roof feeds the tank. A machine-learning model trained on the simulations predicted the required tank within 25% for 97.0% of facilities in spatial cross-validation, against 8.3% for the dry-period rule of thumb. With a 20% margin its tanks met the target at 97.5% of facilities, but at only 74.5% when the nearest training data were at least 110 km away, where the rule's larger tanks were more dependable. Reliability rose slightly over 1981 to 2025, with a significant trend at 15.1% of facilities. A retrospective seasonal analysis found useful skill only for September to November, coming from tank storage and recent rain rather than El Niño or Indian Ocean Dipole indices. The results are screening estimates that need confirmation on site; the code and data are released openly.