Abstract
The automatic extraction of man-made objects from remote sensing imagery is critical for applications ranging from urban planning to disaster relief. This paper proposes an automated algorithm that integrates a graph model with the manifold ranking algorithm. Initially, a priori values for man-made objects are estimated using symmetric and contrast features. A graph model is then established to represent spatial relationships among pre-segmented superpixels, which serve as graph nodes. Multiple characteristics—including color, texture, and main direction—are used to compute the weights of adjacent nodes. Manifold ranking is applied to explore the relationships among all nodes in the feature space, generating a ranking map that indicates the likelihood of an object being man-made. Experimental results demonstrate that the proposed algorithm achieves a high recognition rate and low omission rate compared to traditional segmentation methods.
Introduction
Mapping human infrastructure—buildings, roads, and bridges—from satellite and aerial imagery is a cornerstone of modern geospatial analysis. However, the sheer complexity of remote sensing data, with its varied textures and environmental noise, makes "automatic" extraction a persistent challenge. Researchers from the Satellite Surveying and Mapping Application Center in Beijing have developed a two-stage framework that combines geometric intuition with semi-supervised graph learning to address these hurdles.
The Challenge of Automation
Extracting man-made objects usually falls into two categories: feature-based and model-based. Feature-based methods analyze low-level pixels but often struggle with noise and lack global context. Model-based methods use predefined templates, which are precise but brittle when encountering non-standard structures. While modern deep learning offers high performance, it requires extensive labeled datasets. This research seeks a "middle way": a feature-based approach that leverages global spatial relationships through manifold ranking to improve accuracy without the need for massive manual annotations.
Proposed Methodology
The proposed algorithm operates in two distinct stages: generating a priori map to identify "seeds" and then applying manifold ranking to propagate these labels across the image.
1. Generating the Priori Map
Man-made objects typically exhibit high contrast with their surroundings and high internal symmetry.
- Contrast: The method utilizes KSW entropy to identify high-contrast regions, with thresholds optimized via a Quantum Genetic Algorithm (QGA) for efficiency.
- Symmetry: Multiscale superpixels are used to identify "maximal discs"—geometric areas showing strong internal symmetry. These features are fused with a vegetation mask to create a "priori map," which serves as the initial set of queries for the graph model.
2. Graph-Based Manifold Ranking
The image is over-segmented into superpixels using the SLIC algorithm. These superpixels become nodes in a graph, described by their LAB color, Gabor and LBP textures, and gradient direction.
Using the seeds from the priori map, the manifold ranking algorithm explores the relationships between all nodes. It effectively "spreads" the probability of being a man-made object from the seeds to similar-looking, adjacent superpixels, ensuring structural continuity even if parts of a building were missed in the initial scan.
Experimental Results
The algorithm was tested on 20 remote sensing images, including samples from the UCMerced LandUse Dataset. These images featured complex scenes such as freeways, intersections, and dense urban buildings with high resolution (0.1 m).
The researchers compared their method against two industry standards: the Markov Random Field (MRF) model and the active contour approach.
- Detection Percentage (DP): The proposed method reached 94.8%, significantly higher than MRF (73.7%) and active contour (75.4%).
- Branch Factor (BF): Measuring the ratio of incorrectly detected objects, the proposed method achieved a low 0.068, compared to 0.095 and 0.133 for the other methods.
Qualitatively, the algorithm was superior at "cleaning up" shadows and texture noise that typically confuse other models, leading to much cleaner binary maps of urban infrastructure.
Conclusion
By integrating geometric priors with graph-based manifold ranking, this approach provides a robust solution for automated object extraction. Future work will focus on expanding the feature set and refining the fusion logic to handle even more diverse land cover types, such as distinguishing between symmetric natural formations and human-made structures.
Summary updated