Introduction
Detecting changes in Synthetic Aperture Radar (SAR) images—such as moving aircraft or new construction—is a significant challenge in remote sensing. SAR sensors are inherently prone to speckle noise and perspective distortions caused by variations in flight attitude and sensor angles. Traditional pixel-based or feature-based methods often struggle with high false alarm rates and computational complexity. This paper proposes a "knowledge-aided" approach that leverages a pre-constructed site model to simplify registration and focus detection on specific Regions of Interest (ROI).
Methodology
The proposed framework operates in three distinct stages:
1. Site Model Construction
An offline binary site model is created using high-resolution reference images or CAD data. This model identifies ROIs, such as runways, and "special regions," such as bodies of water, which appear as smooth, dark surfaces in SAR imagery. These regions serve as reliable landmarks for subsequent processing.
2. Geometric Topological Registration
Rather than performing exhaustive pixel-by-pixel matching, the algorithm aligns images using the centroids of the identified special regions. It assumes a similarity transformation (scaling, rotation, and translation). To optimize the matching process, the authors employ an area-ratio heuristic ($S_{i'} / S_i = c^2$), where $c$ is the scaling factor. This allows for rapid pruning of incorrect point-pair candidates, significantly speeding up registration compared to iterative methods like ICP.
3. Target Detection and Analysis
Once registered, the algorithm restricts its search to predefined ROIs. A Constant False Alarm Rate (CFAR) detector is used to identify potential targets, followed by clustering and area-based filtering to isolate aircraft. Finally, shape matrix moments—rotation-invariant features—are used to compare targets across images, determining if they have moved, appeared, or disappeared.
Experimental Results
The method was evaluated using multitemporal SAR images of an airport. By focusing on the runway ROI, the system successfully categorized the status of three distinct airplane targets:
- Target 1: Identified as "No change."
- Target 2: Identified as having "Moved."
- Target 3: Identified as "Disappeared."
The use of prior knowledge allowed the system to ignore background noise and focus exclusively on relevant changes, drastically reducing false alarms and computational overhead.
Conclusion
The site model-based approach offers a robust solution for SAR change detection by reducing the search space and simplifying the registration process. While the current implementation requires manual initialization of the site model and assumes rigid transformations, it demonstrates high accuracy in identifying target movements in complex, noisy environments. Future work will focus on automating model creation and extending the method to handle non-rigid distortions.
Summary updated