Slum mapping in imbalanced remote sensing datasets using transfer learned deep features

Unprecedented urbanization, particularly in countries of the Global South, results in the formation of slums. Here, remote sensing has proven to be an extremely valuable and effective tool for mapping slums. Recent advances in transferring deep features learned in fully convolutional networks (FCN) allow the specific structural types and alignments of buildings in slums to be mapped. The class imbalance of slums is especially challenging in the context of intra-urban variability of slums themselves, and their possible similarity to other urban built-up structures. Thus, in our study we aim to analyze the transfer learning capabilities of FCNs for slum mapping with respect to training on imbalanced datasets and the quantity of available training images. When the slum sample proportion is increased an improvement of the Intersection over Union (IU) of 10% to 30% can be observed. Increasing the total number of images improves the IU up to 20% to 50%. Transfer learning proves extremely valuable in retrieving information on complex and heterogeneous urban structures such as slum patches.