Abstract
This paper presents a method for reshaping and extracting markers and masks of dense houses from Digital Surface Models (DSMs) based on mathematical morphology (MM). In high-density housing areas, buildings in a DSM often appear merged, making standard segmentation difficult. We propose a workflow to first label building markers and then segment them into masks using a watershed transform. To ensure accurate marker detection, the DSM is morphologically reshaped into "domes" using disk-shaped structuring elements sized according to house scales estimated via granulometry.
Introduction
Accurately extracting individual building footprints is a cornerstone of urban planning and 3D city modeling. While LiDAR offers high precision, photogrammetry-derived DSMs are often preferred for their cost-effectiveness and broad coverage. However, these DSMs frequently suffer from a "merging" effect in dense residential districts, where houses are so close that standard algorithms fail to separate them. This paper addresses the challenge of extracting individual house footprints from such "stuck-together" data using advanced morphological operations.
Methodology
The proposed pipeline prioritizes geometric shape over spectral data to overcome the limitations of traditional segmentation.
Granulometry for Scale Detection
Before processing, the algorithm must determine the size of the target objects. The authors use granulometry—a morphological histogram—to detect the scales of off-terrain objects. By applying a sequence of "openings-by-reconstruction" with increasing structuring element (SE) sizes, a Pattern Spectrum (PS) is generated. A sharp peak or step change in this spectrum indicates the dominant size of houses in the area.
Dome Reshaping and Marker Detection
The core innovation is "dome reshaping." The DSM is processed through iterative differential openings to "round" building corners and slightly shrink footprints in the data. This creates distinct "domes" for each house, effectively carving artificial "valleys" between previously touching structures. These domes serve as reliable markers for a marker-controlled watershed segmentation, which defines the final house masks on the original DSM gradient.
Key Results
The method was evaluated using 0.5-meter resolution DSMs of high-density residential suburbs in Sapporo, Japan.
The "Reconstructed Domes" (Figure 4b) demonstrate the algorithm's ability to create isolated spots for each house, a stark improvement over the raw NDSM (Figure 4a) where buildings appear as a connected grid. In primary test areas, the authors reported that all houses were correctly labeled and separated. Even in extremely dense areas, the algorithm maintained high accuracy, successfully extracting individual buildings where standard methods would have merged them.
Conclusion
By treating the DSM as a 3D geometric surface and applying advanced morphological operations like dome reshaping, this method successfully extracts individual footprints in challenging urban environments. Future work will focus on integrating more aggressive vegetation filtering and developing automated 3D polygon modeling from the extracted masks.
Summary updated