Abstract
Traditional building change detection methods often neglect elevation information, leading to inaccuracies in complex urban environments. This paper proposes a novel algorithm that fuses multi-temporal LiDAR data with ortho images (DOMs) for 3D building change detection. By combining height and spectral information through thematic segmentation, the method leverages both pixel-level differencing and object-level analysis. Experimental results on a 1 km² area in Guangzhou, China, demonstrate that this fusion approach achieves 81.8% completeness and 85.7% correctness for buildings larger than 80 m², marking a 10% improvement over image-only methods.
The Challenge of Urban Change Detection
Accurate detection of urban changes is vital for planning and disaster management. However, traditional 2D spectral analysis faces significant hurdles:
- Spectral Confusion: Buildings, roads, and bare ground often share similar spectral signatures, making them difficult to distinguish.
- Geometric Distortion: Shadows and perspective tilts in aerial imagery can create false change signals.
While Digital Surface Models (DSMs) from stereo imagery offer some relief, they are frequently noisy. This research utilizes LiDAR to provide precise 3D height data, effectively separating buildings from their flat surroundings.
A Hybrid 3D Detection Pipeline
The proposed method employs a multi-stage pipeline that integrates pixel-level change detection with object-level classification.
1. Data Preparation and Differencing
Two periods of LiDAR data are filtered and interpolated to generate corresponding DSMs. A "differential DSM" (dDSM) is then created by subtracting the two models, highlighting potential height changes.
2. Filtering and Segmentation
To eliminate noise from moving objects (e.g., cars) or vegetation, the dDSM is processed using height thresholds (2.5m), texture entropy filtering, and morphological operations. The resulting binary change map is then used for thematic segmentation in eCognition, grouping pixels into meaningful "objects."
3. Feature Fusion and Classification
The core of the detection logic lies in the synergy between height and color. The researchers found that combining Mean Height (MH) from LiDAR with the Ratio of Green Band (RG) from ortho images is the most effective way to distinguish buildings from trees and ground. A Decision Tree classifier then categorizes these objects as 'Newly-built', 'Heightened', or 'Demolished'.
Key Results and Insights
The fusion of LiDAR and ortho imagery significantly outperforms single-source methods:
- Accuracy Boost: The system achieved a 10% improvement in both completeness and correctness compared to using ortho images alone.
- Vertical Change Detection: Unlike 2D methods, this 3D approach successfully identifies "Heightened" buildings—structures where floors were added without changing the footprint.
- Object-Level Advantage: By analyzing objects rather than individual pixels, the method avoids the "salt-and-pepper" noise typical of pixel-based change detection.
Conclusion
This research demonstrates that combining LiDAR-derived height information with spectral data from ortho images provides a robust solution for 3D building change detection. While the method excels for buildings larger than 80 m², future work will focus on improving the detection of smaller structures and minor building appendages, which are currently often filtered out as noise.
Summary updated