Using Deep Convolutional Neural Networks for the Identification of Informal Settlements to Improve a Sustainable Development in Urban Environments

Kurzfassung

Currently about one-quarter of the world’s urban population live in slums. Slums are defined by the United Nations (UN) as informal settlements or areas deprived of access to water, sanitation and durable housing. The buildings in slums are overcrowded and lack land tenure security. Slum-identification studies are very much driven by the persistence and growth of slums and the emergence of new slums being inexorably part of contemporary urbanization processes, particularly in the global south where rapid slum development is linked to the failure of formal land markets and low planning capacity. Identifying slums is an import aspect in urban environments of mega-cities. The information on location, boundaries and population in informal settlements is of great need for social economic studies and thus providing beneficial insight for a sustainable urban development. Beyond the identification of informal settlements and their physical parameters it is of great interest to provide these areas with an optimal fresh water-pipe infrastructure, since their supply of water is very limited. The view from above using remote sensing data makes it possible to grasp the physical spatial settlement structures and, accordingly, to approach the characterizing parameters of slums and with this in mind image class segmentation on slum mapping can be done using different approaches. In recent years mainly object based, machine learning and texture classification approaches have been used to identify slums in urban areas. Regular machine learning tasks are limited because of their manually designed features. Another disadvantage of those methods is the inability to transfer the classifier to different datasets. This study provides a combination of methods in deep learning to achieve respectable accuracies in mapping informal settlements. Detected slums provide the prerequisite for establishing an optimal water supply network for all informal settlements. Since this procedure depends very much on the input geo-data, multiple ways of slum mapping using deep convolutional neural networks are presented and the cost of an optimal water-pipe network supplying all slum dwellers with water is calculated for Mumbai and Delhi. Class segmentation performance was evaluated using overall and class based accuracy metrics. Using a pre-trained fully convolutional network resulted in an overall Pixel Accuracy for informal settlements of 78% and a mean Intersection over Union of 68%, while fine-tuned FCNs could achieve an overall Pixel Accuracy for informal settlements of 75% and a mean intersection over union of 63%. Using the best performing FCN a water supply infrastructure was built optimized to the shortest path connecting all slums using different approaches. The investment of a pipeline network providing clean water would cost about 16 million e for Mumbai and 12 million e for Delhi after 10 years of operation.