AI Takes on Extreme Climate:
Global Early Warning and Disaster Response in Practice

Traditional numerical weather prediction (NWP) solves atmospheric dynamics equation systems, a path pursued for nearly 70 years at extremely high cost—the European Centre for Medium-Range Weather Forecasts (ECMWF) consumes roughly 8,400 kWh of electricity for a single high-resolution deterministic forecast run, with per-run costs ranging from €1,000–€20,000. Data-driven "weather foundation models" offer a fundamentally different path: instead of solving equations, they directly learn the statistical patterns of atmospheric evolution from historical reanalysis data dating back to 1979.

Four landmark models

GraphCast (Google DeepMind, Science 2023): a graph neural network deterministic model at 0.25° resolution, completing a 10-day forecast in under 1 minute, outperforming ECMWF HRES on most metrics. GenCast (Nature 2024-12): upgraded to a diffusion-model probabilistic forecast system, generating a 15-day global ensemble forecast on a single TPU v5 in about 8 minutes, outperforming ECMWF ENS on 97.2% of 1,320 verification targets, with the advantage ratio for lead times beyond 36 hours rising to 99.8%. Microsoft Aurora (Nature 2025): positioned as an "Earth system foundation model," capable of forecasting weather/air quality/ocean waves/tropical cyclone tracks. NVIDIA Earth-2/CorrDiff: specifically designed for "downscaling," roughly 500× faster and about 10,000× more energy-efficient than traditional high-resolution numerical forecasting.

Worth noting: the NOAA National Hurricane Center officially incorporated AI models into its tropical cyclone forecasting operational workflow for the first time in the 2025 hurricane season—AI weather models have transitioned from "auxiliary validation tools" to "a link in the decision chain."

Floods are among the natural disasters causing the broadest economic losses, exceeding $40 billion globally each year, and AI's coverage speed in this domain is the fastest of all extreme weather applications.

Google Flood Hub has been under development since 2017, initially relying on scattered hydrological station data from various countries before shifting to a globally unified LSTM model. As of 2025–2026, it covers over 150 countries and roughly 2 billion people, providing nearly 250,000 forecast points, and uses "virtual gauge" technology to cover data-sparse regions with no ground observations at all. River flood warnings can be issued up to 7 days in advance; urban inundation warnings added in 2025 can provide 24-hour advance notice—the training approach is quite representative: using Gemini to analyze decades of public news reports, identifying and structuring over 2.6 million historical flood events across 150+ countries, which in turn train the warning models. Microsoft SAR Flood Model, based on 10 years of synthetic aperture radar data, can penetrate clouds and detect floods at night, identifying nearly 3× more high-risk areas in Ethiopia than existing datasets.

Beyond warnings: distributing money before floods occur

More noteworthy than the warnings themselves is a new form of humanitarian aid catalyzed by AI—"anticipatory action": distributing cash directly to households likely to be affected, based on predictions, before floods actually occur. From June–October 2024, GiveDirectly and IRC, funded by Google.org, distributed cash assistance to 3,000 households in Adamawa State, Nigeria, 5–7 days before flood peaks arrived—Africa's first AI-driven anticipatory cash relief project.

The core contradiction in wildfire detection is time—from ignition to loss of control often takes only tens of minutes, while traditional satellites may take hours to revisit the same location. "Satellite constellations + AI" is compressing this time window from "hour-level" to "minute-level."

FireSat / Earth Fire Alliance was jointly initiated by Google Research, Muon Space, and the Environmental Defense Fund. The long-term goal is to deploy over 50 satellites forming a constellation, theoretically capable of detecting fire spots as small as 5×5 meters with global updates approximately every 20 minutes—but this is a early 2030s long-term design target; by the end of 2026 it will only provide early-adopter institutions with at least twice-daily data. The first satellite, "Protoflight," launched in March 2025, and the first three operational satellites launched in July 2026, marking the system's entry into "initial operational capability." NOAA Next Generation Fire System (NGFS) fuses GOES satellite data with AI, providing initial detection of 19 ignition points during the March 2025 Oklahoma large-scale fires, estimated to have helped avoid over $850 million in structural property damage—250× the system's development cost (under $3 million).

On the commercial ecosystem side, San Francisco startup Pano AI uses a camera + computer vision solution deployed across the western US, Australia, and Canada, with cumulative funding of approximately $81 million (reports note it still relies heavily on manual review); generative AI (GAN/VAE) was first used in 2024 for California and Southern European wildfire scenarios, capable of generating thousands of fire spread simulation scenarios in seconds to assist firefighting resource dispatch.

Compared to the visual impact of floods and wildfires, droughts and heat waves are "chronic" disasters whose destructive power is often underestimated, yet AI deployment in this area is equally intensive. Microsoft AI for Good and Catholic Relief Services used machine learning to predict food insecurity in southern Malawi up to 4 months in advance with 83% accuracy. 2024 was the hottest year on record—according to a Lancet report, India's heat waves caused approximately 247 billion potential labor hours lost and about $194 billion in economic losses.

Typhoons and landslides

Three years of NOAA HAFS testing data show an 8% improvement in track forecasting and 10% in intensity forecasting, with particular strength in predicting typhoon "rapid intensification." Huawei Pangu Weather reduced the prediction error for typhoon eye positions at 3-day and 5-day lead times by approximately 26% and 28% respectively compared to the ECMWF high-resolution system; it became operational at the Hong Kong Observatory in October 2023, compressing typhoon track prediction time from 5 hours under traditional methods to 10 seconds. For landslides, a team at IIT Mandi developed a low-cost sensor + AI solution that can provide up to 3 hours of advance landslide warning with over 90% accuracy, deployed at over 60 high-risk points in Himachal Pradesh; after the 2024 Hualien earthquake in Taiwan, CNN and Vision Transformer-based tools identified 7,090 landslide locations covering over 75 km² in approximately 3 hours.

AI's value in disaster management is not limited to the "prediction" link alone; it permeates the complete chain from pre-disaster to post-disaster. Post-disaster loss assessment is currently one of the links with the clearest AI ROI: the CLARKE system developed by a Texas A&M University team uses drone imagery + AI to assess building and road damage in a 2,000-household community within 7 minutes; Microsoft AI for Good completed damage assessment within 4 hours after the Lahaina fire in Hawaii (over 100 fatalities, losses exceeding $6 billion), with 97% accuracy, directly providing disaster relief maps to the American Red Cross.

The United Nations and the global map

In 2022, UN Secretary-General Guterres launched the "Early Warnings for All (EW4All)" initiative, targeting universal coverage by 2027. According to the WMO "Global Status of Multi-Hazard Early Warning Systems 2025 Report," 119 countries (60%) globally have established multi-hazard early warning systems, a 113% increase from 2015—but only 43% of small island developing states possess such systems, representing the most prominent gap. A comprehensive early warning system can reduce disaster mortality by at least 6×, and 24-hour advance warnings can reduce losses by up to 30%.

China's "Mazu (MAZU)" solution deserves separate mention: Multi-hazard, Alert, Zero-gap, Universal, integrating Fengyun satellites and models such as "Fenglei," "Fengqing," and "Fengshun," it is the world's first national-level action plan responding to the UN Early Warnings for All initiative, reportedly announced to be driving deployment in 30 countries.

Google (DeepMind weather models, Flood Hub, FireSat), Microsoft (Aurora, AI for Good, SAR flood model), NVIDIA (Earth-2/CorrDiff), and Huawei (Pangu) constitute the technology base suppliers in this field. The real capital driving force lies in the insurance industry: according to Aon's "2025 Climate and Catastrophe Insight," global natural disaster economic losses in 2024 were approximately $368 billion, of which only $145 billion was insured, leaving a protection gap as high as 60%; Hurricane Helene topped single-disaster losses at $75 billion.

This massive gap, combined with the real-world pressure of State Farm declining to renew approximately 72,000 California policies starting in 2024 due to wildfire risk, has driven intensive fundraising for AI-driven catastrophe modeling companies: ZestyAI (property-level wildfire risk modeling), CAPE Analytics (property risk assessment), ICEYE (satellite flood/fire monitoring, $65 million Series E completed December 2024). According to a ZestyAI survey of approximately 220 insurance executives, 68% believe AI helps more effectively manage climate-related losses. On the startup ecosystem side, Pano AI, FloodMapp, Floodbase, Satellites on Fire, and Pula (Kenyan agricultural parametric insurance) together form an active vertical startup ecosystem.

AI's progress in extreme climate scenarios is real, but accompanied by several unresolved and rather fundamental problems.

Systematic underestimation of extreme events is the most consistent criticism from academia. Research published in Science Advances points out that the physics-based HRES model "consistently outperforms state-of-the-art AI models such as GraphCast, Pangu Weather, and Fuxi in predicting record-breaking extreme events," with AI models generally underestimating the intensity of high temperature, low temperature, and strong wind records—the fundamental reason being that extreme events are inherently rare samples in training data, and training methods targeting mean squared error inherently tend to generate "over-smoothed" prediction fields, thereby flattening extreme peaks.

• ▸Interpretability ("black box") problem—the prediction process is difficult to decompose, posing potential misleading risks once it influences major public decisions
• ▸Data inequality—AI climate case studies in Africa and South America are significantly fewer than in Asia and Europe, precisely in the regions with the highest disaster vulnerability and the scarcest observational data
• ▸The irony of compute and energy consumption—inference is power-efficient, but training-stage energy consumption is staggering (each Pangu Weather sub-model requires 192 GPUs training for 16 days)
• ▸Vendor self-reported data requires independent verification—"99.9% detection accuracy," "20-minute global refresh," and similar claims are mostly design targets or prospective modeling and should be treated as metrics pending verification

Looking at the progress of 2024–2026 together, AI has moved from a marginal tool in climate science to a core link in national-level disaster prevention systems—it can do more things, faster, and cheaper, but it has not, and is unlikely to in the foreseeable future, fully replace traditional numerical weather prediction and ground observation networks.

The truly pragmatic path is a hybrid "AI + traditional numerical models + human judgment" system: incorporating models like GenCast and Pangu as additional ensemble members into existing forecasting workflows, while retaining physics-based models like HRES/ENS as fallback safeguards for extreme events. For developing countries and resource-constrained institutions, the more realistic recommendation is to prioritize leveraging existing international public goods—Google Flood Hub, China's "Mazu" solution, CEMS's GloFAS/GWIS, etc., most of which are free or low-cost and open, directly bypassing the bottleneck of sparse local observation networks.