Addressing Underrepresentation in Urban Data: Evaluating a Sentinel-2 Super-Resolution Approach for Mapping Informal Settlements in South America

Kurzfassung

Super-resolution describes a technique for upscaling images from lower to higher resolution. This has proven successful for various applications, ranging from medical image analysis to building segmentation in remote sensing. However, the application of super-resolution to highly complex and dense surfaces, such as informal settlements, has not been widely examined. This is particularly relevant because very-high-resolution (VHR) data are temporally, spatially, and financially limited, especially in informal settlements. Applying openly available Sentinel-2 (S2) imagery could therefore improve data accessibility for applications such as disaster risk management, population estimation, and the visibility of informal settlements, which are often excluded from official maps. This thesis focuses on four cities in South America with morphologically heterogeneous informal settlements: Medellín (Colombia), Rio de Janeiro (Brazil), São Paulo (Brazil), and Santiago (Chile). It uses a super-resolution approach by Panangian & Bittner (2025) built on the Generative Adversarial Network (GAN) architecture, but including location embeddings. Both a baseline model trained on USA data and a fine-tuned model for informal settlements in South America are compared. After assessing the super-resolved imagery against reference VHR and S2 data, informal settlement boundaries are evaluated using a segmentation model with a U-Net ResNet18 architecture. The study further employs Monte Carlo Dropout to estimate prediction uncertainty and a leave-one-city-out approach to assess the generalizability of the segmentation model. The results indicate that the baseline super-resolution approach is not applicable to the study areas. The generated outputs contain hallucinated geographical features, which strongly limit the generalizability across different geographical contexts. The fine-tuned approach shows potential for reconstructing both highly dense informal and more structured formal areas. However, in some cities, hallucinated small white buildings were observed, prompting the need for further research. Statistically, the fine-tuned approach increased IoU by 45%, precision by 33.6%, and recall by 31.6% in relation to the baseline model. Relative to the S2 model, however, the enhancements of the fine-tuned model were only minor, with increases by 2.1% in IoU, 0.0% in precision, and 5.0% in recall. Overall, substantial between-city variation was observed, confirming the highly heterogeneous morphologies of informal settlements. In conclusion, the approach shows considerable potential. Further research is needed to address hallucinations, increase the generalizability of both the super-resolution and segmentation models, and investigate and refine the approach’s applicability to small, short-term, highly dynamic informal settlements, such as those in Santiago, in greater depth.