Publications

Salinity intrusion has rendered groundwater undrinkable for over 35 million people in coastal Bangladesh, forcing women and girls to walk hours for fresh water during the dry season. We evaluate a BRAC-led programme that provided household rainwater harvesting systems — 2,000-litre tanks with ceramic filtration, installed on rooftops or with purpose-built catchment structures where rooftop installation was not feasible — to vulnerable households in Mongla Upazila, using a difference-in-differences design with household fixed effects. The intervention substantially improves water access, lowers collection burdens, and reduces water costs during the monsoon and post-monsoon period. However, fixed tank capacity is exhausted before the dry season ends, leaving a majority of treatment households still water-insecure outside the monsoon window. Rainwater harvesting is a viable but calendar-bounded adaptation strategy — the monsoon can be caught, but not stretched.

Can satellite images and low-cost machine learning predict where a river delta will build new land, accurately enough to plan where people settle? We test this on the Lower Meghna Estuary in Bangladesh, turning 35 years of Landsat imagery (1991-2025) into annual maps of where water sits. A Random Forest and a ConvLSTM neural network are each scored against the honest benchmark of assuming nothing changes year to year. Both beat it, but barely: the Random Forest cuts forecast error by 9.0%, the ConvLSTM by 1.9%, and the two cannot be told apart. Standard image-similarity scores miss this entirely: copying last year's map unchanged already scores 0.911. For 2030 the forecast identifies 54,309 hectares of new land, 55.0% of it east of 91.0°E, the only stretch where the models have real skill; in the western channels they are no better than a coin flip. Reported area swings 23-fold with the analysis grid chosen, a distortion the native 100-meter resolution removes. Tested against what actually happened in 2020-2025, the total area is accurate to within a tenth in the best of three trials and a third in the worst, but only 0.349 of flagged parcels became land. That makes this a tool for deciding where to send survey teams, not for allocating plots. A simple tide-line map locates new land just as well; the models earn their place by sorting parcels into tiers that verify in the predicted order, which the tide map cannot do. We set this out as Stability-Led Infrastructure Development (SLID), a phased settlement protocol that reports how often each tier actually verified.