Hong Kong has more than its fair share of adverse weather. Meteorologists are working around the clock to ensure that perilous conditions like severe thunderstorms and deluges can be more accurately predicted by studying cloud formations. Amber Wu says artificial intelligence technology has changed the game in rainfall forecasts.

Hong Kong meteorologists and computer scientists are teaching artificial intelligence models to read rainfall signals hidden in clouds — hours before the clouds start to form.

Researchers, led by experts from the Hong Kong University of Science and Technology (HKUST), have unveiled a storm forecasting system powered by a diffusion model. The deep generative AI model can make weather predictions 15 percent more accurate than the current 30 to 40 percent baseline by simulating and interpreting how clouds develop at a spatial scale of 48 kilometers.

READ MORE: HKUST-developed AI model 'able to forecast storms 4 hrs ahead'

The system, which is being tested by the China Meteorological Administration and the Hong Kong Observatory, extends the lead time for predicting dangerous convective storms, including thunderstorms and extreme rainfall, from the current maximum of two hours to four hours.

From cloud formation to rain

The scientists fed the model cloud imagery through China’s FY-4A meteorological satellite and captured data every 15 minutes from 2018 through 2021. The data help the model to simulate cloud evolution from formation to growth, dissipation, merging or splitting.

The model is also acquired through data from remote sensing experts’ image processing, such as cloud-top temperatures, along with human wisdom, including the expertise of frontline weather forecasters and specialists from the National Satellite Meteorological Center.

Low cloud-top temperatures often indicate that a cloud has grown to high altitudes, typically signaling severe convective weather. Meanwhile, forecasters, drawing on years of experience, can pinpoint where rain is forming in cloud images.

With such comprehensive training, once the model receives a new cloud image, it can quickly show not only how that cloud would evolve over the next four hours in 15-minute steps, but also whether those changes would bring rain, explains Su Hui, chair professor at the HKUST’s Department of Civil and Environmental Engineering, who led the research.

Such an ability to predict clouds has significantly extended the weather forecast window.

According to Su, weather forecasts in Hong Kong and globally rely primarily on ground-based radar, which emits microwave signals into the atmosphere and analyzes the echoes reflected back by raindrops, ice crystals and other particles. Human forecasters then determine the size and concentration of those particles, helping them to estimate rainfall accurately.

However, raindrops and ice crystals reflect echoes only after they have formed in the sky. But, this method, Su says, isn’t truly “forecasting”. Satellite cloud imagery works differently by making predictions as air rises and cloud droplets start forming without waiting for them to coalesce into raindrops.

Meteorologists have been making cloud imagery-based weather predictions since the 1960s, but highly variable conditions have made it impossible to reliably translate cloud images into rainfall forecasts based on experience alone until AI changed the game.

Being able to process massive amounts of data, AI models can identify patterns across years or even decades of information. Coupled with powerful generative models like diffusion prototypes producing high-quality images, this forecasting method has advanced dramatically.

However, Su says the model still has limitations — the biggest being that it currently can’t forecast specific rainfall amounts as ground-based radars can because the rainfall volume depends on factors like a cloud’s inner structure.

Her team will continue refining the model by possibly shrinking the minimum observation scale from four kilometers to 100 meters or even tens of meters. To further improve the accuracy of probable rainfall predictions and eventually the volume, the meteorologists intend to introduce more diverse data into the model’s training. This could include more extreme weather event data, lightning and even atmospheric wind fields, says Su.

A black cat in a dark room

Despite rapid technological advances, weather forecasting continues to face considerable challenges.

Leung Wing-mo, former assistant director of the Hong Kong Observatory, once likened predicting typhoons to “searching for a black cat in a dark room” — an analogy that Chen Fei, associate head at the HKUST’s Division of Environment and Sustainability, calls a vivid snapshot of today’s climate prediction struggles.

Chen — a seasoned meteorologist with decades of urban climatology experience at institutions, including the World Meteorological Organization and the United States’ National Center for Atmospheric Research — says current forecasting models struggle in cities like Hong Kong.

The special administrative region lies in a typhoon-and-monsoon-active cyclone belt where localized heavy rain can develop rapidly in a high humidity environment. What makes it more complicated is the city’s intimidating density of high-rise buildings. The concrete forest is highly prone to the urban heat island effect that significantly affects cloud formation and rain intensity. Hong Kong’s mountainous terrain and long coastline also mean that extreme weather can vary sharply by district.

All these factors have made predicting weather particularly difficult — a challenge exacerbated by climate change — says Chen.

Loaded with more energy because of global warming, the atmosphere holds more clouds and moisture — key drivers behind Hong Kong’s extreme weather in recent years. According to the Hong Kong Observatory, last year saw the city break multiple records in temperatures, typhoons and black rainstorm warnings. From late July, Hong Kong hoisted its highest black rainstorm warning five times in a month. Over the year, it also issued 14 tropical cyclone warnings — more than double the long-term average of about six annually, and the highest yearly count since 1946, including two maximum No 10 alerts.

Despite these difficulties and challenges, Chen believes the Observatory deserves praise for its performance, noting that the forecaster is considered to be professional not just in the Asia-Pacific region, but globally with a solid reputation, particularly in typhoon monitoring and public communication.

Impact-based warnings

Hong Kong can now deliver relatively accurate quantitative forecasts at a specific location one to three hours ahead for small-scale, convection-driven extreme precipitation, buying precious preparation time for the public and government agencies.

For further improvement, while continuing to extend the forecast lead time — a task faced worldwide — Chen expects Hong Kong’s weather forecasts to move beyond simply delivering weather information to impact-based warnings. Rather than just reporting wind strength or rainfall volume, public warnings should also incorporate the weather’s potential economic and social effects, collaboration among sectors, and how communities should get involved in prevention work.

For example, a heavy rain warning needs to remind residents of low-lying areas and significant effects, and advise them to avoid specific areas. Chen says such warnings will gradually connect and mobilize the entire society, including meteorological agencies, emergency services and local communities before disaster strikes, boosting a city’s ability to withstand extreme weather. Such ability is known in the industry as climate resilience.

Strong climate resilience saves lives and prevents large-scale power outages, traffic paralysis and other social breakdowns, enabling the city to rebound quickly following a weather disaster, Chen says, recalling the September 2023 “black rain” that broke Hong Kong’s rainfall record dating back to 1884. That deluge flooded underground shopping malls and the Wong Tai Sin MTR station, and paralyzed public transportation. Yet, Hong Kong bounced back almost overnight in a very impressive way, says Chen.

At an international symposium on extreme rainfall held in the SAR in August last year, international experts shared the same observation. “They think we’re moving in the right direction in climate adaptation,” Chen says. “And, they see Hong Kong as having strong climate resilience.”

Resilience in action

In recent years, climate resilience has become a growing topic in public and academic discourse.

In December 2025, the State Key Laboratory of Climate Resilience for Coastal Cities — one of Hong Kong’s 15 State Key Laboratories managed by the nation’s Ministry of Science and Technology — was launched. Co-established by the HKUST and the Hong Kong Polytechnic University, the laboratory aims to enhance infrastructure resilience in Hong Kong and coastal cities worldwide. Along with the other 14 labs, it receives annual funding of up to HK$20 million ($2.55 million) from the HKSAR government, underscoring the city’s commitment to strengthening climate resilience.

Charles Ng Wang-wai, director of the state key laboratory for climate resilience and a veteran geotechnical engineer, says developing climate resilience calls for a complete service chain ending with the public. Sensitive forecasting systems represent the chain’s upstream end, while the downstream end includes robust support spanning rescue, medical services and, most critically, a city’s interconnected infrastructure that underpins every aspect of daily life.

He cites two disasters that had been directly triggered by extreme weather — the deluge in July 2021 in Zhengzhou, Henan province, which triggered severe flooding, killing nearly 400 people, and the collapse of Guangdong province’s Meizhou-Dabu highway in May 2024 following prolonged heavy rainfall that claimed 52 lives. Behind what the lab calls such “compound climate disasters” is infrastructure that fails to withstand extreme weather, including public transit, drainage, power supply, bridges, roads and geotechnical maintenance.

In Hong Kong, key infrastructure areas also include slopes, which Ng considers to be among the most critical, coastline maintenance and wave-resistant structures because the city faces frequent typhoons and heavy rainfall, with over 60 percent of its total land area having natural slopes. “These are the city’s blood vessels,” says Ng. “We must keep them flowing.”

Citing a current project in his laboratory, Ng says scientists are exploring nature-based solutions to replace concrete retaining walls to stabilize slopes against landslides. The facility has also developed a flexible barrier that’s easier to install and far more eco-friendly, producing much less carbon dioxide than concrete. Researchers are studying how to further strengthen plants’ water absorption power to reduce rainfall infiltration, protect slopes and reinforce them by building globally leading physical and numerical simulation models.

While Hong Kong’s academic strength in related fields, particularly geotechnical engineering, is formidable, challenges remain, Ng warns.

He notes that the SAR’s limited experimental space means that many large-scale tests need to be conducted on the Chinese mainland. One such example is the country’s largest outdoor flume facility — 190 meters long and six meters wide — built in Kunming, Yunnan province, in 2024 in collaboration with the Chinese Academy of Sciences. Moreover, in today’s AI era, Hong Kong lacks enough experimental data to leverage the technology for empowering infrastructure development. He calls for data sharing, including slope collapse data, within the Guangdong-Hong Kong-Macao Greater Bay Area, saying this is a key direction for the lab.

He insists that the findings will provide scientific evidence for policymakers that can be translated into action in a timely manner and benefit the public, including residents and the construction sector.

In its 2026 outlook issued in March, the Hong Kong Observatory warned that four to seven tropical cyclones are expected to be within 500 kilometers of the city — a “near-normal” level. Annual rainfall should also be near normal, but there could still be non-uniform distributions of rainstorms and heavy rain.

With a possible El Nino effect — a meteorological phenomenon that often brings wetter springs and winters to Hong Kong — developing in the second half of the year, along with climate warming, Hong Kong’s annual mean temperature is expected to be above normal, with a good chance of being among the 10 warmest years on record.

The challenges never go away.

Ng describes climate resilience succinctly: Regardless of the climate changes, a city must keep running. “Let’s keep this ‘unkillable’ spirit alive.”

Contact the writer at amberwu@chinadailyhk.com