TL;DR
This paper introduces a novel method for updating building maps in urban areas by combining object extraction from aerial imagery with building height estimation. This hybrid approach effectively detects newly built, demolished, and reconstructed buildings, significantly improving map accuracy and reducing errors compared to existing methods.
What Problem Does This Paper Address?
Keeping building maps up-to-date in rapidly changing urban environments is a critical but challenging task. Traditional methods often rely on comparing aerial imagery from different time points, but this requires having data from previous epochs, which isn't always available. When only existing 2D building maps and new aerial imagery (orthoimages and Digital Surface Models - DSMs) are available, accurately detecting changes like new constructions, demolitions, or significant reconstructions becomes difficult, especially in dense urban areas with complex structures and potential occlusions from vegetation. The paper aims to bridge this gap by developing a robust and automated system for building map updates using readily available data.
Methods
The proposed approach integrates two main components:
- Building Object Extraction for Newly-Built Detection:
- A Normalized DSM (NDSM) is generated by processing the DSM to remove ground features.
- Large buildings are identified by thresholding the NDSM.
- Small buildings with gabled roofs are detected using a Local Surface Normal Angle Transform (LSNAT) method, which analyzes roof plane orientations.
- Small buildings with plane roofs are extracted using a Marker Controlled Watershed Segmentation (MCWS) method, which is an improved watershed segmentation technique that uses markers to avoid over-segmentation and accurately delineate object boundaries.
- The final extracted building objects are then compared with existing 2D building maps. Newly identified building objects in areas without prior building polygons are flagged as newly-built.
- Building Height Estimation for Demolished/Reconstructed Detection:
- The local ground altitude surrounding each building is estimated by finding the minimum height within an "outskirt buffer" of the existing building polygon.
- The building altitude is estimated by finding the minimum height within an "inner polygonal buffer" inside the existing building polygon, accounting for potential DSM inaccuracies at building edges.
- The building height is calculated by subtracting the building altitude from the ground altitude.
- By comparing the estimated building height to a pre-defined threshold, buildings that are significantly shorter than expected are identified as demolished or reconstructed.
Finally, the results from both detection methods are integrated to update the building map.
Datasets
The paper evaluates its method on a test area in Tokyo, Japan. The data used includes:
- Existing 2D building vector maps from 2004.
- Newly acquired aerial orthoimages (0.2m resolution) and DSMs (0.5m resolution) from 2008.
The paper does not introduce new benchmark datasets but rather uses existing data for evaluation.
Key Insights and Ideas
- Height as a Key Indicator: The paper highlights that building height is a crucial feature for detecting changes like demolitions and reconstructions, especially when prior imagery is unavailable.
- Hybrid Approach Synergy: Combining object extraction from imagery with height estimation from DSMs and existing maps creates a more robust system than either method alone.
- Addressing Dense Urban Challenges: The MCWS method is specifically designed to handle over-segmentation issues common in dense urban environments with noisy gradient images.
- Practical Applicability: The technique has already been integrated into a practical system for building map updating, demonstrating its real-world utility.
- Beyond Change Detection: The method can also identify misregistrations in existing building maps, serving as an accuracy check.
Key Results
The proposed method was tested on a dataset in Tokyo, Japan, comparing its output to reference data derived from image interpretation.
- Overall Detection Rate: 88.6% of changed buildings were detected out of 44 changed buildings.
- Newly-Built Buildings: 9 out of 13 newly-built buildings were detected (approximately 69% detection rate). Some smaller ones were missed, potentially due to vegetation.
- Demolished Buildings: All 16 demolished buildings were detected (100% detection rate).
- Reconstructed Buildings: 14 out of 15 reconstructed buildings were detected (approximately 93% detection rate).
- The approach demonstrated effectiveness in reducing misdetection and false detection, particularly for demolished and reconstructed buildings.
Limitations and Future Work
- Vegetation Influence: The paper acknowledges that vegetation can obscure newly-built buildings, leading to missed detections. Further improvement in vegetation removal or handling is needed.
- Small Building Detection: While the MCWS method improves small building detection, the paper notes that very small buildings can still be challenging to extract accurately.
- Future Work: The authors suggest further improvements in building object extraction, particularly for very small buildings that are susceptible to vegetation interference.
Summary updated