MACHINE VISION FOR URBAN MORPHOLOGY [SUMMARY]
In Collaboration with Cagri Hakan Zaman (MIT, Computer Science and Artificial Intelligence Laboratory)
Proliferation of data centric methods in mapping practices brings about the question of whether they can integrate the urban morphology and its implications on spatial data analysis. While quantitative data are processed within geographic information systems (GIS) framework through an extensive set of spatial data extraction and processing tools, qualitative assessment of city form is mostly a manual task that requires careful examination of formal and material features. In this research, we introduce a novel computational method for the analysis of the morphological features of cities ranging from macro scale street networks to building stock patterns. We use image processing and pattern matching techniques that are often used in computer vision algorithms for the assessment of morphological features that of street networks, parcel and building stock. Through a comparative analysis of four neighborhoods in Istanbul, we show that there is a strong parallelism between socio- economic development and urban form of Istanbul.
We apply two steps of image processing in order to analyze building stock features. In the first step, the input map was processed to allow boundary detection (Figure 1). A basic color filter was applied to map to reveal the street network. A canny edge detector was applied to the filter image for extracting parcel boundaries in pixel domain. Following the edge detection, each pixel boundary was traced for converting the image data into a list of polygon coordinates of parcels.
Calculating a measure of regularity requires identifying texture features over the building stock image calculated in the previous step. We compute gray level co-occurrence matrix (GLCM), which has been commonly utilized for detecting texture features. It allows finding recurring intensity values in gray scale images, and generate statistics as to the correlation of values, energy, and contrast of repeating patterns. In the context of texture detection various statistical features are defined in relation to GLCM. Here we are interested mostly in correlation. GLCM is defined as:
GLCM is symmetrically calculated in both vertical and horizontal directions. The calculation window that defines the width of correlation matrix determined as w=400, that corresponds to roughly 50 meters in length in real-world dimensions of map features.
We calculate the grid level of a street network based on the shape features of the parcels that are surrounded by the network. Basically, rectangularity of a parcel determines how much the street network is similar to a grid. Therefore, grid level of the street network is calculated following the steps in the previous section: detecting the parcel boundaries and calculating their orientation, and rotating the parcel to its major orientation, resulting as the parcel polygon P. After applying rotation, a bounding box, B, is calculated which is defined as the smallest rectangle that covers the polygon. Rectangularity is then calculated as the ratio between P and B ( g= P/B).
Grid levels of street layouts in 4 case study areas are visualized as heat maps. Sultangazi has the most grid-like street pattern, which is followed by Bahcelievler, Zeytinburnu and Kagithane in order. Morphology maps of street network align with the historical formation of each district. Zeytinburnu and Gultepe were established as squatter settlements by rural-urban migrants between the 1950s and 1980s, thus their urban structure is self-organized and exhibits an exceptionally higher level of complexity. On the other hand, Sultangazi and Bahcelievler, which are planned residential districts, have grid-like layouts. Comparing Sultangazi and Bahcelievler, street pattern of Sultangazi is more regular than Bahcelievler because Sultangazi was established on a vacant land that was not restricted by major road network. However, in Bahcelievler, the map shows that grid-like urban structure of the district turns into irregular patterns close to the highway and its rink road. Comparing Zeytinburnu and Kagithane, both of the maps show dominant irregular layouts and clusters of grid-like structures that are distributed randomly.
In parallel with regularity analysis of street network, we created maps based on parcel size and orientation. In Zeytinburnu and Kagithane, the size and orientation of parcels differs greatly, which creates another layer of complexity besides irregular street networks. Maps show that average sizes of parcels in newer districts are higher than the ones in older districts.
Morphological maps based on grid-like level of street network:
Morphological maps based on orientation of parcels:
Morphological maps based on area size of parcels: