Locating Informal Urban Settlements Bob Bell 1 , Rajesh Veeraraghavan 2 1 Georgetown University 2 Georgetown University {bob.bell, rajesh.veera}@georgetown.edu Abstract The main idea of the paper is that convolutional neural networks can be applied to very high- resolution satellite imagery in order to classify New Delhi into formal (planned colony) vs. informal settlements (Jhuggi Jhopri Clusters). We show that very high-resolution satellite imagery along with convolutional neural networks can achieve high classification accuracy of 95.81%. We find that pre- trained deep learning models for computer vision trained on standard image datasets can be effec- tive for classification of informal settlements using satellite imagery, even when there is not a signifi- cant amount of training data. Deep learning mod- els can learn image features without hand-crafted features and when coupled with the proliferation of cloud-based computer vision services could de- mocratize the analysis of satellite imagery for hu- manitarian and developmental purposes. 1 Introduction By 2030, Indian cities are projected to be home to over 630 million residents but there are clear signs that governance ca- pacity and patterns of public investment are inadequate 1 . The role of data and technology in making cities ”smart” have gotten a lot of attention, but they are typically focused on the wealthy. India’s government has proposed a nationwide pro- gram to build 100 smart cities that focus on such problems of the wealthy and the US has committed significant resources to supporting this initiative. At the same time, Rapid urban- ization around the world and in India, have lead to the rise of mega cities, where the rural poor are increasingly migrating to cities in large numbers. Cites and governments are unable to cope with even basic information about where the poor live to adequately provide them with basic services. We don’t know much about how the poor live as they migrate to the cities. The costs of doing surveys remain expensive, and it is partic- ularly hard for resource starved countries in the global south to keep update the information about where the poor are. Yet, there is a promising research direction, where studies are us- ing satellite imagery using night lights to find aggregate mea- 1 www.un.org/ sures of where the poor live. There is a promising path set out by work done by Blumentstock that combine big data analy- sis with surveys, which point to a productive possibilities to estimates of poverty [Blumenstock, 2016]. This paper contin- ues in that path, with satellite based work to understand how the poor are living. Further, because of the underdeveloped statistical capacity in many countries in the global south, especially for these ur- ban informal settlements, it is important to provide alternative and complementary indicators for helping local governments, civil society, identify segments of urban areas and track pub- lic services accordingly. The Indian government defines a smart city as a city “equipped with basic infrastructure to provide a decent qual- ity of life, and a clean and sustainable environment through the application of some smart solutions” 2 . It is estimated that over the next few years India’s infrastructure investment will be in the $ 1.5 to $ 2 trillion range. While smart cities have gotten a lot of recent attention in a city like Delhi, the discussion has been dominated by understanding ques- tions of planned settlements, there has been relatively very little that we know about the informal settlements in Delhi. The planned settlements in Delhi occupy only a quarter of Delhi’s population, yet these plans of using data and tech- nology reflect the global priorities where data amplifies ex- isting inequalities [Toyama, 2011][Eubanks, 2018]. A pri- mary reason for the increased attention in planned settle- ments is that there is more information available on the for- mer and we don’t even have an adequate picture of the infor- mal settlements[Bhan and Jana, 2015]. There has been as- sessments based on qualitative studies that show that despite being the best resourced city in India, three quarters of Delhi’s population live in “unplanned settlements” that have limited or no access to basic services such as piped water and san- itation, few if any public facilities and are poorly connected to the city’s transport grid [Bhan and Jana, 2015]. Yet, by its own admission, the Indian government and city governments have a very limited understanding of the spatial distribution of basic infrastructure in the poor parts of the Delhi and its socio-economic effects [Heller et al., 2015]. There is a nascent quantitative literature on urban India that has drawn attention to the spatial dynamics of exclu- 2 http://smartcities.gov.in/content