Abstract
Identifying the precise address and identity of buildings in satellite imagery remains a significant challenge for geospatial intelligence. This paper presents a framework that models building identification as a Constraint Satisfaction Problem (CSP). By integrating satellite imagery, vector maps, and non-traditional sources like online phone books, the system applies urban planning logic to accurately label structures even with incomplete or noisy data.
Introduction
In geospatial intelligence, knowing the physical location of a building is often insufficient; one must also know its identity and address. The 1999 accidental bombing of the Chinese Embassy in Belgrade serves as a tragic reminder of the consequences of outdated or disconnected geospatial data. The core technical hurdle is data integration: satellite imagery, road maps, and residential directories often exist in silos and do not align perfectly. This research proposes an automated reasoning approach to bridge these gaps.
Methodology: The CSP Framework
The authors treat building identification as a logic puzzle, specifically a Constraint Satisfaction Problem (CSP). The environment is modeled using variables and a set of common-sense urban planning constraints.
Variables and Constraints
For each building detected in an image, the system assigns two primary variables: the street name ($s_i$) and the address number ($a_i$). The solver then applies several constraints:
- Parity: Even and odd numbers typically reside on opposite sides of a street.
- Ordering: Address numbers generally follow a monotonic increase or decrease along a street.
- Existence: Addresses found in external directories (like phone books) must correspond to a physical building identified in the imagery.
The system utilizes CPlan, a constraint satisfaction planner, to explore valid assignments that satisfy all rules simultaneously.
Experimental Setup
The approach was validated using two datasets:
- Synthetic Dataset: A controlled layout of 29 houses and 4 streets designed to test robustness against missing data.
- Real-World Dataset: A residential block in El Segundo, CA, containing 34 houses. This test utilized satellite imagery from Terraservice, NGA vector data, and residential records from Whitepages.com and property tax databases.
Results and Analysis
The system demonstrates high resilience to incomplete information. In synthetic trials, the framework maintained high recall even when significant portions of the phone book data were removed.
In real-world testing on the El Segundo block:
- Phone Book Data: Achieved 94.1% recall but only 54.7% precision due to the inherent incompleteness of public directories.
- Property Tax Data: When provided with a complete and accurate database, the system achieved perfect 100% precision and recall.
Discussion and Future Work
A key insight is that global constraints allow the system to "fill in the blanks" for missing data points. However, the current model relies on "hard" constraints, which may fail in regions with irregular numbering. Future research aims to implement Probabilistic CSPs to handle soft constraints and improve scalability for city-wide applications. Additionally, the system could be expanded to include region-specific rules for international numbering systems.
By framing map-making as a logical inference problem, this approach provides a robust blueprint for automated, self-verifying geospatial intelligence.
Summary updated