Spaces:

---

ibm-esa-geospatial

/

challenge

like 38

Running

Howdy 🤠, for this Blue Sky Challenge submission, we find ourselves in the outskirts of Houston, Texas. The project aims to explore the concept of surban growth through the lens of TerraMind. Specifically, developing a new kind of framework for urban planners and real estate developers to understand where a city has been and where it's going, and who might be on a similar journey.

Data Preparation
The Dynamic World Dataset was queried from January 1st, 2015 to December 31st, 2025 for 38 Houston suburbs. These LULC images were aggregated to an annual level via mode. The Dynamic World Dataset land cover classifications were mapped to their ESRI LULC Dataset equivalents in order to be compatible with TerraMind. These images were then divided into 224x224 tiles and embeddings were generated using TerraMind V1 base. The patch level embeddings were aggregated to the image tile level via mean, ensuring to mask out sentinel embeddings resulting from patches that were majority no data. As metadata, the percentage of each land cover classification per tile was calculated. To create an embedding at the suburb level the suburb’s tile level embeddings were aggregated via mean.

The Latent Timeline of Suburbia

Now with a 10 year time series of embedded LULC maps per suburb, each one can be treated as an object moving through a 768-dimensional space. I designed metrics that summarized these suburbs' trajectories in the embedding space over time. The summarization came down to 3 metrics:

• Average velocity (Rate of Semantic Change): This measures the average magnitude of annual change in a suburb's embedding vector. It identifies the "speed" of urban evolution, distinguishing between suburbs undergoing rapid, high-impact redevelopment and those with a slow, incremental evolution.

• Net Displacement (Total Regime Shift): This metric calculates the Euclidean distance between a suburb’s initial state in 2015 and its final state in 2025. A high net displacement indicates a fundamental "Regime Shift" in the suburb’s character.

• Path Efficiency (Developmental Intent): Calculated as the ratio of Net Displacement to Total Path Length, this metric quantifies the "directness" of a suburb's evolution.
  -- High Efficiency (~1.0) suggests a deliberate, linear developmental trajectory toward a specific urban archetype.
  --Low Efficiency (<0.5) indicates "semantic vibration," where a suburb is undergoing erratic or non-linear change, such as seasonal fluctuations or uncoordinated infill redevelopment.

Then the K-nearest neighbors algorithm was used to cluster the suburbs on these three metrics. With the following results:

The clusters can be characterized as follows:

• Cluster 0: The Big Shifters
  These suburbs have the highest average velocity and total displacement. They are undergoing large changes where the very character of the suburb is being rewritten over the decade.

• Cluster 1: The Balanced Developers
  These show moderate velocity and high path efficiency. They are moving purposefully in a specific direction.

• Cluster 2: High Churn Zones
  They have high annual change (velocity) but very low efficiency, meaning they are changing a lot on the surface without actually moving toward a new identity.

• Cluster 3: The Outliers
  This cluster captures the most extreme developmental pivots where both velocity and displacement are peaked.

The same logic was applied on the percentage of each LULC class for each suburb. Those outputs are below:

A lot of times LULC percentage changes are what urban planners have to work with, so it is an interesting baseline to compare to. Especially as we dive into the case studies.

Case Studies
Fulshear is consistently mentioned when searching which are Houston's fastest growing suburbs. It has had a population increase of 236% between 2018 and 2023. When looking at its LULC map, there is definite indication of urban expansion. Its embedding trajectory is in cluster one which indicates a balanced trajectory to urbanization. This validates the method against a known growth narrative.

The Cloverleaf case study is a bit of a puzzle. It is a suburb that has one of the highest displacements in the embedding space but one of the lowest displacements in the LULC percentage space. Perhaps this indicates a “Hidden Regime Shift” whereby change is detected that traditional LULC maps cannot capture.

Conclusion
Ultimately, this work explores what becomes possible when land-use maps are enhanced by embedding them into the TerraMind model. By following suburbs through TerraMind’s embedding space over time, it becomes feasible to ask comparative questions—which places are developing in similar ways, which are stabilizing, and which are quietly changing character—that are difficult to answer using LULC percentages alone.

The results suggest that embeddings provide additional signal: they capture shifts in urban form and function that may occur even when class proportions remain largely unchanged. For investors, planners, or analysts, this offers a way to reason about development trajectories using more than surface-level land-cover change, helping to contextualize growth patterns that otherwise appear similar on a map.

TerraMind supports not only the mapping of cities, but also reasoning about their future trajectories, complementing traditional LULC representations of urban change.

Code: https://drive.google.com/drive/folders/1aHCBJnK_e4w_ys3U21QMgzgqURxvyj0X?usp=drive_link
Email: erikagutierrez2921@gmail.com