The study presents a single generative foundation model as a potential way to broaden the use of optical and radar data for disaster assessment, environmental monitoring and urban analysis. Optical satellite imagery records sunlight reflected from Earth’s surface, so thick clouds can obscure the ground and nighttime conditions prevent clear imaging. Synthetic aperture radar, or SAR, uses transmitted radio waves and the signals returned from the surface, allowing it to acquire images at night and in cloudy weather.

SAR observations differ in brightness and pattern from ordinary photographs, making them difficult to interpret. GeoCore-9B was designed to process both types of satellite data through representations learned specifically from overhead Earth imagery. The reported results concern image restoration and conversion in public-dataset evaluations; they do not establish that the model can already forecast ground changes or operate as a comprehensive disaster-assessment system.

At the center of GeoCore-9B is a generative image model with about 9 billion adjustable internal parameters, the numerical values through which a neural network represents and generates patterns. Rather than adapting an image generator originally developed for ordinary photographs, the KAIST team trained the core component from scratch on Earth-observation data. Its training material included terrain, forests, ports and cities, allowing the model to learn the visual characteristics of different land surfaces and urban environments before it was adapted to satellite-image tasks.

The team classified the result as a foundation model because its pretrained core could serve as the basis for multiple downstream applications. The training dataset contained 10 million pairs of satellite images and sentences describing their contents. A port image could be paired with a statement that ships and warehouses were present, while a rural scene could be accompanied by a description of a road running between fields.

This image-and-text pairing gave the model visual and semantic information about the depicted places. Those learned features were used to train the generative component, allowing knowledge acquired from satellite-image generation to be transferred to the later restoration and conversion tasks. A distinctive feature of the approach is that GeoCore-9B learned not only what an image showed, but also where it was located and the spatial scale it covered.

The researchers supplied latitude, longitude and ground sample distance, or GSD, together with the descriptive sentences. Latitude and longitude identify a position on Earth, while GSD indicates the ground area represented by one image pixel. At a GSD of 1 meter, one pixel corresponds to an area of about 1 meter by 1 meter on the ground.

The geographic and scale conditioning was intended to help the model generate images consistent with a specified location and resolution. Scale matters because objects with similar shapes in satellite views can represent different features depending on the actual area shown. A shape interpreted as a small building in one image might correspond to a large industrial facility at another scale.

By learning with location and GSD information, the team sought to reduce this ambiguity and preserve meaningful relationships between image content, real-world position and spatial extent. The approach differs from a general-purpose image generator because the model was trained from scratch on overhead views rather than adapted from a system built to produce conventional photographs. To represent the organization of terrain and cities more accurately, the team introduced a training technique called Geospatial Semantic Alignment.

Under this method, the generative model refers to features extracted by another AI system specialized in satellite imagery. Those features provide a training reference for structures such as road connections and farmland boundaries. The specialized AI is used only during training, so it does not add computation when GeoCore-9B generates images.

In the researchers’ account, the method links the generative model’s learned image features with representations of spatial structure, while leaving the common core available for subsequent task-specific adaptation. After the broad pretraining stage, the researchers fine-tuned GeoCore-9B separately for cloud removal and radar-to-optical conversion. They used LoRA, or Low-Rank Adaptation, which keeps the core model fixed while training a small number of additional parameters.

This approach reduces the burden of retraining an entire large model from scratch for each task. The resulting system therefore combines a shared satellite-image foundation with relatively lightweight adaptations tailored to two observation problems: restoring optical images obscured by clouds and converting SAR data into optical-style images. In evaluations against existing task-specific models using publicly available satellite datasets, GeoCore-9B produced the strongest reported results for restoring cloud-covered images.

It scored higher on key metrics measuring brightness, color and structural similarity to reference images. In converting radar images into optical images, it also achieved the best results on some metrics assessing image quality and fine detail. The findings indicate that knowledge gained while generating satellite images can be applied to the restoration and conversion of observed images.

The release does not provide the metric names, numerical scores, uncertainty estimates or a complete comparison with every competing model, so the size of the reported differences cannot be independently assessed from the supplied information. The model is not yet a forecasting system for events on the ground, and the team plans to validate its performance across a wider range of sensors and regions. Future work will examine changes such as floods, building damage, wildfires and urban expansion, as well as possible defense applications including surveillance and reconnaissance.

For now, the demonstrated capabilities are satellite-image generation, cloud restoration and radar conversion. The proposed benefits for emergency response and environmental monitoring remain expected applications that will require broader testing, particularly across sensors and geographic regions. The study was led by first author Jeonghyeok Do of KAIST’s Information & Electronics Research Institute, with Professor Munchurl Kim of the School of Electrical Engineering as corresponding author.

High-performance computing resources were provided through the Advanced GPU Utilization Support Program of the Ministry of Science and ICT and the National IT Industry Promotion Agency. The findings were accepted for presentation at NeurIPS 2026, and the paper was posted on arXiv on August 3, 2026.

Subject of Research: A geo-aware generative foundation model for restoring cloud-obscured optical satellite images and converting radar observations into optical-style views.

Article Title: KAIST develops AI that reconstructs cloud-covered ground in satellite images with photo-like clarity

Article References: KAIST develops AI that reconstructs cloud-covered ground in satellite images with photo-like clarity. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: GeoCore-9B, satellite imagery, cloud removal, synthetic aperture radar, Earth observation, generative AI, foundation models, geospatial AI, remote sensing, disaster monitoring, KAIST, develops

The study was led by Gia Barboza-Salerno, an associate professor in Ohio State’s colleges of social work and public health and director of the university’s Investigating Spatial Structures in Urban Environments, or ISSUES, Lab. Her research group specializes in advanced spatial statistics, geospatial artificial intelligence and machine learning to understand how neighborhood environments shape health, safety and social well-being. The impetus for the work came from a point she made at a recent conference that resonated with fellow researchers studying urban greenery and violence prevention. When it comes to the effects of green space on communities, local context matters. It is not as simple as planting a tree and expecting that alone to solve the problem of violence, she argued, because researchers still do not fully understand how, or whether, the local context in which greenery sits interacts with a subsequent reduction in violence.

Technically, the study broke new ground by measuring greenery the way a pedestrian actually experiences it. Previous research on green space and violence has often relied on satellite imagery or aerial assessments of vegetation, which capture tree cover from above but say little about what someone walking down a street can see at eye level. Barboza-Salerno’s team instead turned to Google Street View images and machine learning tools to quantify the amount and types of greenery visible from the street where each death occurred. This street-level, pedestrian perspective had not been applied before to the specific question of greenery near the locations of firearm-related deaths, according to the study, making it a methodological advance in a field where the relationship between vegetation and violent crime has long been recognized but poorly understood.

The case-control design was equally deliberate. The researchers examined 899 firearm homicides and 2,256 natural deaths in Cook County in 2023. For each firearm homicide, they compared the location to natural deaths of people similar in age, sex and race who died nearby, a matching strategy that helps isolate the role of the immediate environment from the characteristics of the people involved. By pairing deaths that occurred in close geographic proximity and among demographically similar individuals, the analysis could ask a focused question: holding the who constant, did the where, specifically the visible greenness of the street, change the odds that a death was a shooting?

The answer, in broad terms, was yes. More visible greenery, greater tree canopy and more green space along walkable streets were linked to lower odds that a death in that location was a firearm-related homicide. This finding aligns with a growing body of research showing that urban green spaces, parks and other areas lined with plants, grass and trees, improve community and individual health outcomes by encouraging physical activity, reducing stress and improving air quality. Green spaces have also been associated with reduced gun violence and other violent crimes in earlier work. But the new study’s most striking result concerns the limits of that protection, and it is here that the research departs from a simple greening prescription.

The benefits of green space were not universal. Tree cover and greenery offered the most protection in less disadvantaged neighborhoods, while in the most disadvantaged neighborhoods, being closer to a park was actually linked to a higher risk of firearm homicide. The researchers attribute this pattern to multidimensional measures of social deprivation, including poverty, structural housing disadvantage, economic inequality, family structure and income, which appear to have a stronger impact on violence risk than a neighborhood’s green spaces themselves. In other words, the social and economic conditions surrounding a park can overwhelm, or even reverse, whatever protective effect the park might otherwise provide.

Barboza-Salerno framed the finding as a double inequity. Green space is very important and critical for violence prevention, she said, but not all green space matters equally. If you lived in a socially advantaged environment, having green space mattered more for reducing violence than in a socially disadvantaged area. Green space should be thought of as a violence-prevention intervention, she argued, yet the green space that does the most protective work tends to be concentrated in more advantaged neighborhoods. Disadvantaged areas get less quality green space, and the green space they do get is less protective without other supports. The result is a compounding disadvantage in which the communities facing the highest burden of firearm violence receive both fewer and weaker environmental buffers against it.

The mechanism behind this reversal, the researchers suggest, lies in maintenance and use. When parks in disadvantaged neighborhoods are not well-maintained or used regularly, they can become more of an eyesore than a protective factor. In neighborhoods with more resources, well-kept parks have a beautification effect that can improve how people feel and strengthen neighbors’ willingness to look out for one another, a dynamic that researchers of urban violence often describe in terms of collective efficacy and informal social control. In the most disadvantaged areas, Barboza-Salerno said, there is a lot of economic disinvestment, and that is the association with violence, not the park per se. The park is not the cause of danger; it is the surrounding disinvestment that turns an asset into a liability.

The study’s authors are careful about what their data can and cannot show. The analysis looked only at environments, not individuals. As Taylor Harrington, a doctoral candidate in public health and a co-author of the study, explained, the team only looked at the location of death for the person in the context of the environment at that moment. The measurement does not encapsulate a person’s whole lived experience or their exposure to green space and deprivation throughout life up until that point. This is a case-control study of places, not a longitudinal study of people, and its findings describe associations between street-level greenness, neighborhood disadvantage and where firearm deaths occurred, not causal guarantees that greening alone will reduce shootings in any given block.

Even so, Barboza-Salerno has seen green space succeed as part of a broader effort. As the lead researcher at the Violence Prevention Collaborative in Boston and Director of Research for the Dudley Street Neighborhood Initiative, she partnered with youth development organizations that employed young people to fix up local parks. The work gave the participants new skills and changed how neighbors perceived them, illustrating how greening projects can double as social interventions. Parks can provide safe places for children to play, socialize and have positive reciprocal interactions. Alongside housing and food security, Barboza-Salerno described green space as one part of a comprehensive community-building approach that needs to be scaled to see effective violence prevention in the United States. If we think carefully about where green spaces are placed and why they make a difference, she said, this could be huge. The study’s message for policymakers is that urban planning is a critical part of public health prevention, and that planting trees without addressing poverty, housing disadvantage and economic inequality is unlikely to deliver the safety benefits that greener, better-resourced neighborhoods already enjoy. Other Ohio State co-authors of the study are Balaji Ramesh and Jeffrey Wing of the College of Public Health.

Subject of Research: Street-level urban greenness, neighborhood deprivation, and firearm-related homicide risk in Cook County, Illinois

Article Title: Green space is linked with fewer gun-related homicides — but not everywhere

Article References: Green space is linked with fewer gun-related homicides — but not everywhere. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: urban green space, firearm homicide, gun violence, neighborhood disadvantage, Google Street View, machine learning, geospatial AI, public health, urban planning, Cook County, tree canopy, violence prevention

The award, part of NSF’s Fire Science Innovations through Research and Education program, runs from September 2026 through August 2029 and brings together a multidisciplinary team from the University at Buffalo, the University of California, Berkeley, Northeastern University and the U.S. Department of Energy’s Pacific Northwest National Laboratory. The study, titled Collaborative Research: FIRE-WUI: REKINDLE: Resilient Engineering through Knowledge Integration for NetworkeD Lifelines and Ecosystem, will focus on communities in the wildland-urban interface, the zone where housing, power lines and other human development meet or intermingle with forests, grasslands and other natural vegetation. These are precisely the places where the destructive potential of wildfire and its aftermath is growing fastest.

As wildfires become more frequent and severe across the American West and beyond, wildland-urban interface communities face escalating risks that extend well beyond the immediate destruction of the fire itself. Burned landscapes lose the vegetation and soil structure that once absorbed rainfall, making them dramatically more prone to debris flows and flash floods during subsequent storms. At the same time, fires and their secondary hazards can inflict significant disruptions on lifeline infrastructure, including power grids, water systems and transportation networks. Current approaches to wildfire resilience planning, however, typically consider these interconnected risks in isolation, which makes it difficult to see how they interact with one another and how people actually respond when multiple hazards strike in sequence.

REKINDLE is designed to close that gap by combining artificial intelligence, infrastructure modeling and social science to assess wildfire-related risks together and to identify concrete ways to bolster public safety and community resilience. Sayanti Mukherjee, PhD, the project’s lead principal investigator and an associate professor in UB’s Department of Industrial and Systems Engineering, explained the central problem the team hopes to solve. People are familiar with wildfire risk, she noted, but what is often overlooked are the additional hazards and cascading effects that can follow. A region affected by wildfire can become more vulnerable to subsequent hazards like debris flows and flash floods, and when those events cut off roads or interrupt power or water service, the impacts can quickly compound for communities already struggling to recover.

Mukherjee, who is also an affiliate faculty member with UB’s Department of Electrical and Computer Engineering and the Institute for Artificial Intelligence and Data Science, is joined by co-principal investigator Susan Spierre Clark, PhD, of UB’s Department of Environment and Sustainability. The broader investigative team includes Auroop R. Ganguly, PhD, of Northeastern University; Marta González, PhD, and Anna Serra-Llobet, PhD, of UC Berkeley; and Andre Coleman, PhD, and Sam Chatterjee, PhD, of Pacific Northwest National Laboratory. This mix of expertise in engineering, environmental science, data science and social systems reflects the project’s core premise: that understanding cascading wildfire disasters requires simultaneously modeling the physical hazards, the engineered networks they damage and the human decisions those damage cascades trigger.

Through REKINDLE, researchers will investigate how wildfires and the hazards they set in motion can compromise key infrastructure, producing disruptions that ripple through the essential services people depend on for daily life and well-being. Clark emphasized that the team wants to understand not only how infrastructure systems are affected, but how those disruptions shape household decisions and community response. Bringing those pieces together, she said, can help communities build resilience, restore critical infrastructure and recover faster from future disasters. The framing matters because infrastructure failures are rarely just technical problems; a closed road can prevent evacuation, a lost power connection can disable water pumping, and a prolonged outage can determine whether a family can remain in or return to a damaged neighborhood.

Artificial intelligence will play a central role in the technical core of the project. The researchers plan to build a physics-informed, geospatial-AI model that fuses Earth observation data with wildfire expertise to generate maps showing where wildfires and related hazards such as debris flows and flash floods are most likely to occur. By anchoring machine learning in physical processes, the team aims to produce hazard predictions that are both data-driven and scientifically interpretable, an increasingly important consideration for tools intended to inform emergency management decisions. In parallel, the researchers will deploy machine learning models to trace how failures could propagate across interconnected power, water and transportation systems, capturing the dependency structures that turn a localized fire damage into a region-wide service disruption.

The human dimension of the cascade will receive equal attention. Using household surveys, focus groups and simulation games, the research team will examine how infrastructure failures affect families’ livelihoods and their ability to recover from cascading disasters. These social science methods are intended to reveal how households weigh risks, make evacuation and recovery decisions, and cope when multiple services fail at once. The team will then use these findings to help determine where investments in stronger infrastructure could deliver the greatest protective impact before a fire occurs, and how recovery efforts should be prioritized in the aftermath. Mukherjee noted that deciding where to invest resources, whether through hardening infrastructure or increasing community engagement and outreach, will be an important part of the project, and that AI and optimization can help assess different strategies and pinpoint where investments would have the most value. That kind of insight, she said, can support emergency managers, infrastructure operators and environmental planners in making more informed decisions for their communities.

To keep the work grounded in real conditions, the researchers will build and test REKINDLE through case studies in the Los Angeles and Montecito regions of California, both of which have experienced major wildfires and devastating post-fire debris flows in recent years. Montecito in particular has become a sobering reference point for cascading wildfire hazards, having seen destructive debris flows sweep through burned terrain after intense rainfall. The research team will collaborate with government agencies, utility providers and community stakeholders throughout the project to help ensure that the tools they develop reflect the practical needs of the people and institutions that will ultimately use them, from county emergency planners deciding where to pre-position resources to utilities weighing which assets to reinforce first.

The project also carries a substantial education and workforce development mission. Researchers will create wildfire-focused educational materials for students ranging from middle school through graduate school, introducing a new generation to the science of wildfire resilience and the related hazards that follow fires. Undergraduate and graduate students will have opportunities to gain hands-on experience through internships with industry partners and national laboratories, building career pathways in a field where demand for expertise continues to grow. The team will make its educational resources and research data publicly available so that other researchers, practitioners and communities can build on the work. If REKINDLE succeeds, its integrated view of fire, flood, infrastructure and human response could reshape how at-risk communities plan for the full arc of a wildfire disaster, not just the day the flames arrive.

Subject of Research: AI-based modeling of cascading wildfire hazards and infrastructure resilience in wildland-urban interface communities

Article Title: University at Buffalo to lead $2 million National Science Foundation project on wildfire resilience

Article References: University at Buffalo to lead $2 million National Science Foundation project on wildfire resilience. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: wildfire, cascading disasters, artificial intelligence, infrastructure resilience, wildland-urban interface, debris flows, flash floods, National Science Foundation, University at Buffalo, geospatial AI, community resilience, lifeline infrastructure