Governments and researchers across South Asia are turning to artificial intelligence to anticipate floods, landslides and other climate hazards.
Pakistan’s observatory
Punjab Chief Minister Maryam Nawaz Sharif shared photographs of construction work on the Punjab Climate Observatory in Lahore on Nov 15. The facility is under construction.It is meant to monitor air quality, water resources and quality, soil conditions and weather, and to use an AI-powered climate intelligence system for forecasting, flood prediction and early warnings. Planned features include research laboratories, an environmental data centre, a meteorological garden and public learning facilities. The AI models, suppliers, data sources and technical architecture have not been disclosed.
Punjab approved the observatory in May 2025. The University of Engineering and Technology, Lahore, submitted a technical report, and on May 30, 2025 the Punjab Planning and Development Board met a team from the Chinese Research Academy of Environmental Sciences.
Cost
The estimated cost was PKR 5 billion in 2025. The 2026-27 development programme lists PKR 14.588 billion, formally approved on Dec 8, 2025. Spending reached PKR 500 million by June 2026. The 2026-27 allocation is PKR 1.5 billion, with PKR 12.588 billion projected for 2027-28.
Existing systems
Punjab already has monitoring. Its air quality dashboard receives data from 59 stations, with plans to reach 100, and added satellite and forecasting features on Sept 8 with the International Centre for Integrated Mountain Development (ICIMOD). Pakistan’s Flood Forecasting Division runs 102 automatic weather stations, 114 manned observatories, 45 telemetric gauge sites, eight Doppler radars and about 130 discharge sites, with forecasts from nowcasts to seasonal outlooks.
Nepal’s research and systems
On Oct 31, the journal npj Natural Hazards published a framework for forecasting district-level impacts of monsoon-triggered landslides in Nepal up to 14 days ahead. It uses terrain, past rainfall and sub-seasonal rainfall forecasts, with records from 2016 to 2022 and testing on the 2023 and 2024 monsoons. It reported an F1 score of at least 0.7. That combines precision and recall and is not 70 percent accuracy.
Nepal has weather stations, a C-band radar, flood warnings, a 1155 information line and the BIPAD disaster portal. Community systems use sensors, sirens and SMS and have been developed with ICIMOD and the Nepal Red Cross Society. Simple systems like these can work where central forecasts cannot reach.
What recent disasters show
ICIMOD’s Oct 13 assessment of the Aug 26 Rasuwa disaster described a rock-and-ice collapse, with up to 40 centimetres of ground deformation seen by satellite radar in the 48 days before. It said SMS alerts went out 25 to 30 minutes after the corridor had been struck, and found gaps in cross-border information sharing.
The Oct 11 Budhi Gandaki flood in Gorkha was linked to a landslide dam without heavy local rain. The 14-day landslide model covers monsoon-triggered landslides and is not shown to be able to predict either event.
Limits of AI
The UN’s Early Warnings for All framework says an effective system needs four parts: knowing the risk, observation and forecasting, warning communication, and preparedness and response. AI mainly helps with the second. Poor data, rare events, false alarms and missed warnings, and weak maintenance and coordination can undermine forecasts.
Pakistan’s 2022 floods affected about 33 million people and killed more than 1,730, with losses above US$30 billion. Nepal’s September 2024 floods and landslides affected about 2.6 million people and killed more than 250.




