Neuronal Network Classifier for Urban Density
The urban datum, as an abstract element, is a spatial / temporal witness of a certain event or characteristic of a place, while the code is the language that helps us read these behaviours. The capture, cleaning, normalisation and analysis, together with the application of Machine/Deep Learning techniques, become contemporary processes required when analyzing a space. Code design as a technique must be ‘democratised’ among architects/urban planners since it allows the creation of critical tools that help us study territorial trends and improve decision- making.
In this context, the Urban Density Classifier using Neural Networks is a supervised Deep Learning model that can predict the number of households per surface from an aerial photograph of any city in the world. The tool breaks an orthophoto into tiles and classifies each one according to its density range. In this way, the model not only identifies the way in which the territory is occupied, but also characterises the intensity of the urban footprint and its compactness degree. Predictive models are built using Convolutional Neural Networks (of three types: basic CNN, VGG16 & VGG19) and are trained with a total of 7,500 labelled images of the city of Madrid.
Three ‘urban tastings’ are generated by using this predictive model – predictions of 500 sq km of Spanish territory that have been processed by the tool. This is aimed at materialising the algorithm and the whole automated process (from prediction, to volumetric definition and construction of the object). In this way, a piece/x-ray is obtained, which shows us how we inhabit and the pressure we exert by living in the territory.