Atlas Urban Index: A VLM-Based Approach for Spatially and Temporally Calibrated Urban Development Monitoring

AI summary

Overview

Research area: Geospatial AI / remote sensing, combining satellite imagery (Sentinel-2) with Vision-Language Models (VLMs) for urban development monitoring.

Technical level: Intermediate. The core idea is described in plain language, but familiarity with remote sensing indices (NDBI, spectral bands like SWIR and NIR) and how VLMs are prompted helps with understanding the calibration discussion.

Scope: The paper introduces the Atlas Urban Index (AUI), a VLM-based scoring mechanism for measuring urban development in standardized ~4.89 km × 4.89 km Geohash 5 (GH5) grid cells, demonstrated qualitatively on two regions of Bangalore.

What This Paper Is About

Traditional indices for detecting built-up areas, such as the Normalized Difference Built-up Index (NDBI), are easily thrown off by atmospheric noise, seasonal variation, and cloud cover, which makes large-scale tracking of urbanization unreliable. The authors propose the Atlas Urban Index (AUI), which asks a Vision-Language Model (OpenAI's GPT-4o-mini) to assign a development score on a 0–10 scale to a satellite image of a region, using reference images and the region's previous image to keep scores consistent across space and time. The goal is a more reliable and stable measure of urban development than pixel-based indices, especially where construction is informal and official records are absent or incomplete.

Key Contributions

1. The Atlas Urban Index (AUI): a metric for urban development computed from Sentinel-2 satellite imagery using Vision-Language Models, producing a 0–10 development score per standardized GH5 spatial unit.
2. Two calibration strategies to fix VLM instability: (i) supplying a curated set of manually selected reference images with assigned AUI scores to calibrate scoring spatially, and (ii) supplying the image from the immediately preceding time period along with its score to calibrate scoring temporally and mitigate cloud-related noise. The paper names the problems these address as spatial miscalibration (a more developed region scoring lower than a less developed one) and temporal miscalibration (a region scoring lower at a later date despite visible growth).
3. A processing pipeline for cloud-reduced time series: collecting Sentinel-2 imagery for the first and third quarters of each year, filtering by cloud cover to keep the least cloudy image, and omitting the second and fourth quarters to yield one representative image every six months, stored as TIFF with all 13 spectral bands and converted to JPEG RGB composites (B4, B3, B2) for the VLM.
4. A qualitative comparison against NDBI on two Bangalore regions — the rapidly developing airport area and the more modestly growing Bannerghatta National Park area — where AUI is reported to align with known urbanization trends while NDBI is reported to fail.

Main Findings

• AUI tracks airport-area development (GH5 tdr70): The AUI for tdr70 rises steadily from 7.2 in early 2016 to 8.2 by 2025. Between 2016 and 2019, when only the basic foundation of the airport is visible, AUI rises from 7.2 to 7.4. Between 2019 and 2022, AUI rises to 7.8, corresponding to completion of Terminal 2, expansion of road networks, emergence of commercial infrastructure, and reduction in agricultural land. Between 2022 and 2025, AUI climbs to 8.2, reflecting dense commercial infrastructure and a mature urban form.
• NDBI fails on the same region: The paper reports that NDBI does not reflect the transformation of tdr70 between 2016 and 2025, with NDBI values plotted on a scale roughly between −5·10⁻² and 0.1.
• AUI captures modest growth near Bannerghatta National Park (GH5 tdr0t): AUI increases from 2.5 in early 2016 to 4.1 by 2025. It rises from 2.5 to 3.5 between 2016 and 2019 (modest growth), from 3.5 to 3.9 between 2019 and 2022 (growth tapers off), and from 3.9 to 4.1 between 2022 and 2025 (no visible change, flat AUI).
• NDBI again fails to reflect change: NDBI for tdr0t over 2016–2025 is plotted on a scale roughly between −0.2 and 0.1 and is reported not to capture the transformation.
• Reference anchors span the full score range: The curated reference set comprises Rajajinagar (GH5 tdr1u, AUI 9–10), Halasuru (GH5 tdr1y, AUI 7–8), Kaggalipura (GH5 tdr0g, AUI 5–6), Hesaraghatta (GH5 tdr4f, AUI 3–4), Jallipalya (GH5 tdp54, AUI 1–2), and a forest region near Koothampalyam on the Tamil Nadu–Karnataka border (GH5 tdp49, AUI 0).
• Evaluation is qualitative only: The paper states that its experiments on Bangalore are qualitative; no quantitative accuracy metrics, error rates, or benchmark scores against NDBI are reported.
• Urbanization framing: The paper cites global urbanization rising from roughly 55% of people living in cities today to nearly 70% by 2050, with much growth concentrated in rapidly developing countries such as India, China, and Nigeria.

Methodology in Plain English

The authors divide a city into standardized grid cells called Geohash 5 (GH5), each roughly 4.89 km × 4.89 km. For each cell they gather Sentinel-2 satellite images from the first and third quarters of each year, then keep only the least cloudy image in each quarter, discarding the other two quarters. This yields one clean representative image every six months, giving ten years of history in the Bangalore demonstration.

Each stored image is a TIFF containing all 13 Sentinel-2 spectral bands; the authors extract just the visible RGB bands (B4, B3, B2) and save them as JPEGs, which are what the Vision-Language Model actually sees.

Scoring is done with GPT-4o-mini, but not by simply asking it for a number. That naive approach produced two failure modes the authors describe: spatial miscalibration, where a more developed region gets a lower score than a less developed one, and temporal miscalibration, where a region's score drops after visible growth (or rises after decline). To fix this, the model receives three inputs at once: the current image of the target region, a set of six manually curated reference images each labeled with a discrete AUI range from 0 up to 9–10, and the same region's image from the immediately prior time period together with its previously assigned score. The reference images anchor the scale in space; the prior image anchors it in time and helps absorb cloud or seasonal artifacts.

For comparison, the authors compute NDBI, defined as (SWIR − NIR) / (SWIR + NIR) using Sentinel-2 band B11 for SWIR and band B8 for NIR, averaged over all pixels of an image to produce a single image-level score. They then compare AUI and NDBI time series in two Bangalore regions: the Kempegowda International Airport area (GH5 tdr70) and the Bannerghatta National Park area (GH5 tdr0t, referred to in one appendix heading as "tdrot").

Why This Matters

Impact on research: The paper suggests that pixel-arithmetic indices such as NDBI — which are sensitive to atmospheric noise, seasonal variation, and cloud cover — can be supplemented or replaced by VLMs that reason holistically about an image. It opens a VLM-based route to long-term urban monitoring, and the paper notes the mechanism is general: it can be applied to other standardized spatial units where high-resolution satellite imagery is available, and paired with alternative VLMs.

Real-world applications:

• Real estate developers using growth measures to guide site selection and monitor projects.
• Retailers and restaurants using urban expansion patterns to identify promising markets and monitor the performance of existing stores.
• Government agencies evaluating urban policy outcomes, allocating resources, and supporting sustainable city planning.
• Independent monitoring of informal construction, where government records are absent or incomplete and independent monitoring is the only viable way to track change.

Industry relevance: The work is presented by authors affiliated with Propheus, and an abridged version is slated for presentation at and publication in the Proceedings of ACM IKDD CODS 2025, indicating commercial as well as academic interest in scalable urban development measurement.

Future Directions

• Integrating additional data sources: The authors suggest incorporating places data from open sources such as Overture, which is routinely updated, to refine the AUI.
• Forecasting: Using the AUI time series to forecast future urban development, for applications in real estate development, urban planning, and retail strategy.
• Supervised learning to replace the VLM: Using the current VLM-based model to generate a large-scale dataset of satellite images with corresponding AUI labels, then training a supervised learning model on that dataset to compute AUI directly, which would simplify computation and offer a more scalable, cost-effective solution.
• Cross-region generalization: The authors note that future applications beyond Bangalore could expose the VLM to regionally diverse examples at similar development levels (for example, a desert city in the Middle East and a tropical city in Southeast Asia) to ensure cross-context consistency; this diversification was not required for the Bangalore demonstration because all reference regions share similar environmental characteristics.

Target Audience

This paper is most useful to geospatial and remote sensing practitioners, urban planners and analysts, and applied machine learning researchers interested in using Vision-Language Models for satellite imagery. It will also interest real estate, retail, and government teams that need monitoring of urban growth where official records are sparse or unreliable. Readers looking for a quantitative benchmark comparison will not find one here: the paper's evaluation is qualitative, and no numeric accuracy, error, or calibration metrics are reported.

Authors’ abstract

We introduce the {\em Atlas Urban Index} (AUI), a metric for measuring urban development computed using Sentinel-2 \citep{spoto2012sentinel2} satellite imagery. Existing approaches, such as the {\em Normalized Difference Built-up Index} (NDBI), often struggle to accurately capture urban development due to factors like atmospheric noise, seasonal variation, and cloud cover. These limitations hinder large-scale monitoring of human development and urbanization. To address these challenges, we propose an approach that leverages {\em Vision-Language Models }(VLMs) to provide a development score for regions. Specifically, we collect a time series of Sentinel-2 images for each region. Then, we further process the images within fixed time windows to get an image with minimal cloud cover, which serves as the representative image for that time window. To ensure consistent scoring, we adopt two strategies: (i) providing the VLM with a curated set of reference images representing different levels of urbanization, and (ii) supplying the most recent past image to both anchor temporal consistency and mitigate cloud-related noise in the current image. Together, these components enable AUI to overcome the challenges of traditional urbanization indices and produce more reliable and stable development scores. Our qualitative experiments on Bangalore suggest that AUI outperforms standard indices such as NDBI.