Satellites Reveal How Wealth Shapes Flood Recovery in Pakistan

When the monsoon rains of 2022 submerged up to a third of Pakistan, they did more than drown villages and farmland; they exposed a fault line running through the country’s disaster response. A new study published in Environmental and Sustainability Indicators has now quantified that fault line with unprecedented spatial precision, combining satellite imagery, radar-derived flood maps, and a machine-learning wealth index to show that the poorest communities in Pakistan not only suffer the most severe flood damage but also recover the slowest—and that a second catastrophic flood in 2025 struck before the recovery from the first was anywhere near complete.

The research team, led by Sughra together with Siriporn Darnkachatarn and Yoshio Kajitani, analysed more than 85,000 grid cells covering roughly 379,000 square kilometres of Pakistani territory. Each cell measures approximately 2.1 kilometres on a side, a resolution fine enough to distinguish one floodplain village from its wealthier neighbour a few kilometres away. The 2022 monsoon, the most destructive flood in the country’s recorded history, affected 33 million people, killed 1,739, and destroyed 1.7 million homes, with combined damages exceeding 30 billion US dollars. Before that disaster had fully healed, the 2025 monsoon delivered a second blow, shifting northward into Punjab and Khyber Pakhtunkhwa, displacing three million people, submerging 1.12 million hectares of crops, and killing more than a thousand people nationwide.

What makes the study methodologically distinctive is its refusal to simply compare flooded and unflooded areas. Such comparisons are notoriously biased, because floods do not strike random locations: water pools in low-lying, agriculturally productive, densely settled plains that differ systematically from the hills and deserts that stay dry. To correct for this, the researchers employed a quasi-experimental technique called propensity score matching. A logistic regression model estimated each grid cell’s probability of flooding based on seven pre-flood characteristics: population density, cropland fraction, built-up fraction, elevation, and three years of nighttime light baselines. Flooded cells were then matched to statistically comparable unflooded controls using an Epanechnikov kernel, which weights nearby controls most heavily and assigns zero weight to any control whose propensity score lies more than one bandwidth away. The difference between what actually happened in flooded cells and what the weighted controls suggest would have happened without flooding—the matched difference—serves as the study’s core estimate, though the authors are careful to describe their results as covariate-adjusted associations rather than causal effects, since pre-flood economic trends were not fully parallel between the groups.

The flood maps themselves came from Sentinel-1 synthetic aperture radar, a satellite sensor that can see through clouds and darkness by measuring how microwave signals bounce off the Earth’s surface. Because open water reflects radar energy away from the sensor, flooded pixels show a sharp drop in vertical-horizontal polarised backscatter relative to a dry-season baseline. The team applied province-specific thresholding, speckle filtering, terrain correction, and masks for permanent water bodies and steep slopes, then validated the resulting maps against independent flood products. A cell was classified as flood-exposed if more than 20 percent of its area was inundated—a threshold chosen to ensure substantive flooding rather than marginal boundary overlap. In 2022 this identified 16,116 exposed cells; in 2025, 18,000.

The immediate impacts were staggering but unevenly visible depending on which satellite indicator was consulted. Vegetation, measured by the Normalised Difference Vegetation Index from MODIS, showed a massive 29-percentage-point loss in flooded cells relative to matched controls in 2022, an effect size so large it registered a Cohen’s d above 1.0. Yet nighttime lights from the VIIRS Day/Night Band, a widely validated proxy for economic activity, registered a decline of less than one percentage point. The authors attribute this discrepancy to a known limitation: in poor, rural, low-luminosity regions, there is little artificial light to lose, so the sensor systematically under-registers agricultural and livelihood damage. An elasticity analysis suggested that nighttime lights miss roughly half a percentage point of flood-related agricultural harm. The lesson for disaster monitoring is clear—no single satellite proxy tells the whole story, and relying on economic luminosity alone would dramatically understate the suffering of precisely the populations most in need.

Recovery trajectories told an equally instructive story. After the initial 2022 shock, flooded areas showed positive cumulative differences in nighttime light growth by 2023 and 2024, a pattern consistent with reconstruction-driven activity in which rebuilding, relief investment, and restored productive capacity temporarily push output above the counterfactual path. But the authors caution that flood-exposed and control cells were already diverging before the flood, so part of this rebound may simply reflect pre-existing trends. Whatever its origin, the recovery was fragile. When the 2025 flood arrived, the year-on-year trajectory of the 2022-flooded cohort turned decisively negative, with event-specific estimates showing a 2.1 to 2.9 percentage-point decline in economic activity growth. Vegetation, by contrast, rebounded quickly through natural regeneration and seasonal cropping cycles—except in arid Balochistan, where recovery-attainment differences remained stubbornly negative at around 13 percentage points, likely reflecting fragile soils and limited irrigation infrastructure.

The study’s most consequential finding concerns wealth. Using Meta’s Relative Wealth Index, which estimates asset wealth at roughly 2.4-kilometre resolution from mobile connectivity, satellite imagery, and topographic features, the researchers stratified all flooded cells into five wealth quintiles. A clear and consistent gradient emerged: poorer cells experienced larger negative impacts in the flood year and slower recovery thereafter. The Spearman rank correlation between wealth and the 2022 impact was +0.167, highly significant, and while it attenuated in later years, the quintile ordering never broke down. The mechanisms are structural: poor households lack access to credit, depend on climate-sensitive agriculture, live in weaker housing, and wield less political influence over reconstruction spending. Compounding the injustice, the spatial analysis showed that the districts with the longest inundation durations in 2022—concentrated in Sindh’s lower Indus basin—also recorded the country’s lowest wealth scores. Physical hazard and socioeconomic vulnerability, in other words, are not independent risks but co-located ones.

Provincial comparisons revealed four distinct recovery typologies. Punjab, with stronger institutions and higher baseline wealth, showed steadily positive economic trajectories reaching +10 percentage points by 2024. Sindh, hit hardest in 2022 with the highest population exposure and the second-highest vegetation loss, suffered a 6.2-percentage-point economic decline and had effectively zero recovery by 2023, remaining negative through 2025. Khyber Pakhtunkhwa displayed a striking divergence, with vegetation recovering strongly while economic activity declined persistently, deepening after the 2025 flood shifted into the province. Balochistan deteriorated late but persistently, recording the largest 2022 vegetation loss of any province at 66 percentage points. Perhaps most alarming, about 12.4 percent of all analysed cells—10,637 of them, concentrated in the lower Indus corridor—were flooded in both 2022 and 2025, and these persistent-exposure zones showed non-linear cumulative damage that exceeded the sum of two isolated events, as coping resources exhausted by the first disaster were unavailable when the second arrived.

The authors translate these findings into concrete policy. They identify six persistently flood-affected, low-wealth districts—Kambar Shahdad Kot, Jacobabad, Kashmore, Larkana, Dadu, and Shikarpur—as priority targets for recovery and resilience investment, and they argue that disaster risk management must explicitly plan for sequential floods, with pre-positioned financing and adaptive infrastructure designed to shorten recovery windows between events. More broadly, the study offers a replicable, near-real-time template for monitoring disaster recovery in data-scarce regions, where official statistics are delayed or unreliable. As climate change increases both the frequency of extreme floods and the probability of repeated exposure in the same places, the effective recovery window will keep shrinking. Without deliberately redistributive recovery policies, the study warns, each new disaster will deepen the inequalities that made the last one so devastating—turning the monsoon from a seasonal rhythm into an engine of entrenched poverty.

Subject of Research: Satellite-based assessment of flood recovery inequality across wealth gradients in Pakistan following the 2022 and 2025 monsoon floods

Article Title: Flood recovery inequality in Pakistan: Evidence from the 2022 and 2025 monsoon floods using satellite data and relative wealth index

Article References: Sughra, Darnkachatarn, S., & Kajitani, Y. (2026). Flood recovery inequality in Pakistan: Evidence from the 2022 and 2025 monsoon floods using satellite data and relative wealth index. Environmental and Sustainability Indicators, 32, Article 101544. https://doi.org/10.1016/j.indic.2026.101544

Image Credits: AI Generated

DOI: 10.1016/j.indic.2026.101544

Keywords: Pakistan, flood recovery, 2022 monsoon floods, 2025 monsoon floods, satellite remote sensing, nighttime lights, NDVI, Sentinel-1, Relative Wealth Index, propensity score matching, climate inequality, Sindh