Roof type classifier

Roof type is a crucial information for wind hazard analyses of buildings because it is a key attribute needed for consideration of wind effects on structures.

There are three major roof types, as shown in Table 2.2.1, that are widely used in the world: flat, gabled, hipped.

Table 1.2.1 Roof prototypes

Fig. 1.2.1 Flat


Fig. 1.2.2 Gabled


Fig. 1.2.3 Hipped

Correspondingly, a typical satellite image of each roof type is shown in Table 1.2.2.

Table 1.2.2 Example satellite images of different roof types

Fig. 1.2.4 Flat


Fig. 1.2.5 Gabled


Fig. 1.2.6 Hipped

Satellite images are a scalable source for inferring roof type information. In the attempt to determine roof type for every building in a region, a ConvNet classifier is trained to take a satellite image of a building and predicts its roof type. A training data set of 6,000 satellite images (2,000 for each roof type: flat, gabled, hipped) is collected. Specifically, ResNet [HZRS16], which is a widely-used ConvNet architecture for image feature recognition, is employed.

The architecture of the model is shown in Fig. 1.2.14. In this module, we used a 50-layer ResNet.


Fig. 1.2.7 ResNet