Washed Away!

By Pulaha Roy and Nandita Banerji

Microsoft's new AI-powered flood dataset highlights India's growing vulnerability, revealing that one in every 10 square kilometres has experienced flooding over the past decade. Globally, the satellite-based analysis shows a steady 5% annual increase in flood extent, a troubling trend with far-reaching implications

South Asia — and India in particular — has emerged as a hotspot for flooding in the last decade. A new analysis by Down To Earth (DTE) of a global flood mapping project led by Microsoft, using artificial intelligence and cloud-penetrating radar satellite imagery, reveals that over 10 per cent of India’s landmass has experienced flooding at least once in the past decade.

The worst-hit states, predictably, include Bihar, Uttar Pradesh, Assam and West Bengal — but leading the charts in raw percentages is Punjab, where 68 per cent of the state’s land showed up as flooded.

At first glance, the figures are startling. But researchers caution that not all floods are the same. Punjab’s unusually high numbers are largely the result of controlled, deliberate inundation of paddy fields, which require standing water through much of the agricultural cycle. But what’s happening in Bihar is a very different story.

In Bihar, the water is far from welcome. More than half the state’s area, 51 per cent, has faced flooding over the past 10 years, much of it from destructive, recurrent river floods. The state’s geography places it at the mercy of the Kosi and Gandak rivers, among others, which frequently burst their banks during the monsoon, displacing thousands and damaging crops, roads and homes.

Uttar Pradesh, India’s most populous state, isn’t far behind, with flooding across 25 per cent of its area. However, in sheer numbers, close to 60,000 square kilometres in the state have been impacted by flooding — the biggest numbers so far. Bihar stood at a little over 48,000 sq km and Punjab saw over 34,000 sq km impacted.

Assam (over 23,000 sq km) and West Bengal (over 31,000 sq km) clocked in at 30 per cent and 37 per cent, respectively. Haryana recorded 38 per cent, with roughly 17,000 sq km impacted.

These figures are drawn from a new global flood dataset developed by Microsoft’s AI for Good Lab, which combines deep learning models with Sentinel-1 radar satellite imagery — a technology that can see through clouds and operate both day and night. The results are a part of a wider study in the journal Nature Communications.

By applying their model to 10 years of satellite data (2014-2024), the team has produced the most comprehensive flood mapping dataset available globally, covering both past floods and areas at continued risk. What makes this dataset especially powerful is its public availability. The flood maps, covering every Sentinel-1 radar image from October 2014 to September 2024, have been released online along with the full codebase used to generate them.

What the researchers found

India is far from alone in facing a rising tide — the study revealed the extent of flooding around the world has been significantly underestimated and Africa has shown some of the sharpest increases in flood-prone areas.

Globally, the model identified 71 per cent more flood-prone land than older methods had captured. In Africa, the detected flood extent nearly doubled; in Ethiopia, it nearly tripled.

The same tool was used in Kenya’s deadly 2024 floods, where it provided real-time flood maps to emergency responders, helping estimate that 75,000 hectares of cropland had been affected; this figure was nearly identical to official ground reports.

By layering flood detection data with land use maps, the researchers turned their attention to agriculture. In Semera, Ethiopia, around 19 per cent of cropland was found to lie in historically flooded zones. That’s nearly three times higher than what previous satellite records, such as Landsat (7 per cent) and MODIS (2 per cent), had shown.

But it was in Dolo Ado, further south along the Ganale River, where the contrast was most striking: more than half — 52 per cent — of cropland was flagged as flood-prone by the new model. Earlier datasets had captured just 1 to 3 per cent. In a region where families depend on rain-fed staples like maize and sorghum, such underestimation carries heavy consequences — both for food security and rural livelihoods.

Bigger, unsettling picture

When the researchers charted flood data across a 10-year timeline, from 2014 to 2024, they spotted what appears to be a slow but steady rise: A 5 per cent annual increase in flood extent. If that trend continues, it could mean a 60 per cent rise each decade — a shift with far-reaching implications.

The map of growing risk is not evenly drawn. A belt stretching from Nigeria to Ethiopia has seen a clear uptick in flooding, in line with projections from global climate models that predict heavier rainfall in parts of Africa. Eastern Australia, too, stands out, reflecting a string of extreme weather events in recent years.

However, the dataset covers just a decade — not quite long enough to draw definitive conclusions about climate-driven change, the authors have warned.