Mapping roofing materials with multispectral satellite imagery and a Convolutional Neural Network as a step toward modeling fire behavior in urban landscapes

This repository contains all code and documentation for mapping roof material types (e.g., shingle, metal, slate, etc.) using multispectral satellite imagery from PlanetLabs SuperDove [ref] trained on a building footprint dataset which combines the Microsoft Building Footprints [ref] and the Zillow Transaction and Assessment Database (ZTRAX) which was used under a University license agreement [ref]. This work was supported by the Rutgers University Open Philanthropy Project, grant # 1433. Refereed manuscript in preparation.

Figure 1. Study area map with (a-b) Planet PSB.SD true-color (Red-Green-Blue) composites, (c-d) log-scaled distribution of roof material classes and (e-f) log-scaled distribution of building indoor areas. AP=Asphalt; CN=Concrete; CS=Composition Shingle; ME=Metal; SH=Shingle; SL=Slate; TL=Tile; TG=Tar and Gravel; UR=Urethane; WS=Wood Shake/shingle.