JODHPUR — In the era of the "Digital Twin," where scientists and urban planners strive to create a perfect virtual replica of our physical world, the accuracy of global datasets is paramount. A recent landmark study led by researchers from the National Remote Sensing Centre (NRSC) of the Indian Space Research Organisation (ISRO) has provided a critical reality check for one of the most ambitious mapping projects ever conceived: the GlobalBuildingAtlas (GBA).
While the research confirms that the GBA is an unprecedented achievement in horizontal mapping—identifying the "footprints" of India’s rapidly expanding cities with near-perfect precision—it uncovers a significant "vertical deficit." The findings suggest that the artificial intelligence (AI) powering this global map systematically underestimates the height of skyscrapers, sometimes by more than 50%. As India stands on the precipice of a massive urban transformation, these findings have profound implications for disaster management, energy planning, and the development of sustainable "smart cities."
Main Facts: A Tale of Two Dimensions
The GlobalBuildingAtlas represents a technological leap forward. It is the first open-access dataset to attempt a three-dimensional mapping of every building on Earth, cataloging over 2.75 billion structures. Previous efforts by tech giants like Microsoft and Google focused primarily on two-dimensional footprints. The GBA, however, aims to provide height estimates and simplified 3D shapes (Level of Detail 1), turning flat maps into volumetric models.
The NRSC study, centered in Jodhpur and analyzing major hubs like Mumbai, Bengaluru, and Hyderabad, reveals a stark contrast in performance:
- Horizontal Precision: In planned residential areas across five major Indian cities, the GBA achieved a "completeness ratio" of over 99%. In cities like Bengaluru and Kolkata, the dataset matched ground reality building-for-building, making it a "goldmine" for tracking urban sprawl and population density.
- Vertical Inaccuracy: The dataset consistently failed to accurately measure the height of tall buildings. The study noted a "systematic negative bias," where the taller the building, the greater the error.
- The Skyscraper Gap: In one of the most striking examples, a 266-meter skyscraper in Mumbai was estimated by the GBA to be just 113 meters tall—a discrepancy of 153 meters.
- AI Training Bias: The researchers identified that the AI models used to generate the GBA were largely trained on architectural data from Europe and North America, leading to misinterpretations of the unique urban densities and building styles found in the Global South.
Chronology: Testing the World’s Newest Map
The testing of the GlobalBuildingAtlas followed a rigorous scientific timeline, utilizing some of the most advanced remote sensing technology currently in orbit.
The process began shortly after the public release of the GBA dataset. Recognizing the need for an independent "stress test," the ISRO-NRSC team selected India as the primary testing ground due to its unprecedented rate of urbanization and the sheer diversity of its architectural landscapes—ranging from dense, informal settlements to ultra-modern glass towers.
To establish a "ground truth" (a reference point of absolute accuracy), the researchers turned to NASA’s ICESat-2 (Ice, Cloud, and Land Elevation Satellite-2). Launched in 2018, ICESat-2 is equipped with the Advanced Topographic Laser Altimeter System (ATLAS). This instrument fires trillions of photons toward Earth; by measuring the time it takes for these photons to bounce back, the satellite can determine the height of objects with centimeter-level precision.
Throughout the study period, the team compared these laser-measured heights against the GBA’s AI-generated estimates for 20 representative structures. Simultaneously, for horizontal verification, the team utilized ISRO’s Bhuvan portal, a sophisticated platform for hosting Indian remote sensing data. They processed high-resolution imagery using the "Segment Anything Model 2" (SAM 2), a cutting-edge AI model, to ensure their reference footprints were as accurate as possible.
The final phase of the study involved synthesizing this data to understand why the GBA succeeded so well in 2D but struggled in 3D, leading to the publication of their findings in mid-August 2026.
Supporting Data: The Magnitude of the Discrepancy
The data gathered by the NRSC provides a granular look at the GBA’s performance across different urban typologies.
Horizontal Success
In the "horizontal domain," the GBA is nothing short of revolutionary. The researchers found that in Bengaluru and Kolkata, the digital map was almost indistinguishable from high-resolution satellite imagery. This 99% accuracy rate suggests that the AI is exceptionally well-trained at identifying the "edges" of human habitation. For land-use studies and population modeling—where knowing where people live is more important than the height of their roofs—the GBA is currently the gold standard.
Vertical Failure
The vertical data, however, tells a story of significant underestimation. The study categorized buildings into low-rise, mid-rise, and high-rise structures:
- Low-rise/Industrial: For warehouses and small residential units, the error margins were relatively low, though still present.
- Mid-rise Buildings: Errors frequently exceeded 40%. A building that was actually 50 meters tall might be represented as 30 meters.
- High-rise/Skyscrapers: This is where the model broke down. The negative bias became exponential. The Mumbai skyscraper case study (underestimated by 57%) highlights a critical failure in the model’s ability to process extreme verticality.
The Technological Bottleneck
The GBA uses "monocular optical imagery." Unlike LiDAR or stereoscopic imagery (which uses two cameras to create depth, similar to human eyes), monocular imagery attempts to guess height from a single 2D photo. The AI looks at shadows, the angle of the building, and perspective cues. However, in dense Indian cities, shadows from one building often fall on another, and the close proximity of structures creates a "cluttered" visual field that confuses the algorithm.
Furthermore, the GBA utilizes "Level of Detail 1" (LOD1). In the world of 3D modeling, LOD1 represents buildings as simple blocks with flat tops. This ignores the complex spires, sloped roofs, and rooftop infrastructure (like water tanks and helipads) that are ubiquitous in Indian cities, further contributing to the height deficit.
Official Responses and Expert Perspectives
While the creators of the GlobalBuildingAtlas have celebrated the dataset as a milestone for open science, the ISRO-NRSC study serves as a formal "cautionary note" from the scientific community.
The researchers emphasize that the GBA’s reliance on AI trained on Western data creates a "geographical bias." Urban forms in the Global South do not always follow the neat, grid-like patterns of New York or London. The dense, multi-functional nature of Indian urbanism requires a more nuanced algorithmic approach.
"For anyone trying to understand a neighborhood’s density or the extent of urban sprawl, the GBA is an exceptionally reliable tool," the study notes. However, it explicitly warns that for applications requiring vertical precision, the data is currently insufficient.
Experts in remote sensing suggest that this study should act as a "roadmap for future improvements." The suggestion is clear: global datasets must integrate more "active" sensing data, such as the laser measurements from ICESat-2, rather than relying solely on "passive" optical imagery and AI guesswork.
Implications: Why Height Matters for the Future
The findings of the ISRO-NRSC team are not merely academic; they have real-world consequences for how we build and protect our cities.
1. Disaster Management and Flood Modeling
One of the primary uses of 3D city models is to predict how water will move through a city during a flood. If a building’s height and volume are underestimated by 50%, the entire hydrological model becomes flawed. Emergency services might underestimate how many floors of a building remain safe during a surge, or how wind patterns between skyscrapers might affect the spread of a fire.
2. Energy and Sustainability
As cities move toward solar energy, planners use 3D models to calculate the "solar potential" of rooftops. If a building is mapped as being much shorter than it is, the shadows it casts on neighboring buildings will be calculated incorrectly, leading to failed investments in solar infrastructure. Furthermore, vertical data is essential for calculating the "Urban Heat Island" effect, as tall buildings trap heat differently than low-rise structures.
3. Smart City Infrastructure
India’s "Smart Cities Mission" relies on precise data for resource management. Accurate 3D models are needed for everything from telecommunications (placing 5G small cells) to urban air mobility (planning drone flight paths). A map that "shrinks" skyscrapers by 150 meters is a significant liability for these future technologies.
4. The Path to a Global "Digital Twin"
The study concludes that while we are closer than ever to a digital replica of the Earth, we are not there yet. The GlobalBuildingAtlas provides a magnificent "foundation," but the "walls and roofs" of our digital world require more work. By identifying these errors, the ISRO team has ensured that the next generation of maps will be more representative of the diverse, soaring reality of the Global South.
As we move forward, the integration of satellite laser altimetry with AI-driven optical mapping promises a future where our digital maps are as tall, as complex, and as accurate as the cities they represent. For now, the message to urban planners is clear: trust the footprint, but verify the height.
