Abstract
Buildings change detection quantifies temporal effects on urban areas for evolution studies or disaster damage assessment. This paper proposes a shape-matching approach for multi-resolution remote sensing images to avoid the artifacts of traditional image resampling and the "salt-and-pepper" noise of pixel-based methods. By extracting building shapes and performing Centroid-Coincident Matching based on Euclidean distance, the methodology identifies new and demolished buildings. The matching process uses the registration Root Mean Square (RMS) error as a threshold to define corresponding objects, ensuring robustness across different pixel sizes.
Introduction
Traditional change detection typically relies on pixel-by-pixel comparisons. However, when dealing with multi-resolution data—satellite images of the same area taken at different times with different sensors (e.g., 0.5m vs 1m resolution)—standard methods struggle. Resampling images to a common grid can distort spectral information and introduce errors. Furthermore, high-resolution images often suffer from the "salt-and-pepper" effect, where individual pixels are misclassified due to local variance like shadows or roof textures. This paper introduces a geometric approach that is robust to resolution differences and alignment errors.
Methodology
The proposed workflow shifts the focus from pixels to geometric objects through three main steps:
Object-Based Image Analysis (OBIA)
Instead of classifying pixels, the system groups homogeneous pixels into "objects" using eCognition software. These objects are classified into Buildings, Vegetation, and Other. This object-oriented approach significantly reduces noise compared to traditional pixel-based classification.
Polygonal Approximation
To clean up the jagged edges of extracted building shapes, the Douglas-Peucker algorithm is applied. This simplifies the polygons, making them more suitable for geometric comparison while preserving their essential structural characteristics.
Centroid-Coincident Matching
This is the core innovation of the research. For every building polygon in the first image ($T_0$), the system calculates the Euclidean distance to the centroids of all polygons in the second image ($T_1$).
The matching logic uses the Root Mean Square (RMS) error from the initial image registration as a threshold:
- If the distance between two centroids is $\leq$ RMS, the building is considered "unchanged."
- If a building in $T_0$ has no match within that distance in $T_1$, it is marked as Demolished.
- If a building in $T_1$ has no match in $T_0$, it is marked as New.
Experimental Setup and Results
The study utilized a multi-temporal dataset of the ZengCheng district in Guangzhou, China:
- Image 1 (2014): WorldView-2 data at 0.5m resolution.
- Image 2 (2015): Yaogan-24 data at 1m resolution.
The proposed shape-matching method outperformed traditional post-classification comparison (which requires resampling):
- Overall Accuracy: Increased from 94% to 97%.
- Kappa Coefficient: Improved from 0.89 to 0.94.
Visual analysis confirmed that the resulting change maps were much cleaner, successfully identifying distinct new and demolished buildings without the scattered noise typical of pixel-based methods.
Conclusion
Shape matching proves to be a superior choice for multi-sensor studies because geometric centroids remain relatively stable even when pixel values change due to resolution differences. By working in vector space, the method avoids data degradation from resampling. While currently limited to detecting complete additions or removals rather than partial modifications, this approach provides a robust framework for urban evolution monitoring.
Summary updated