Integrating Remote Sensing and Street View Imagery with Deep Learning for Urban Slum Mapping: A Case Study from Bandung City

In pursuit of the Sustainable Development Goals (SDGs)’s objective of eliminating slum cities, the government of Indonesia has initiated a survey-based slum mapping program. Unfortunately, recent observations have highlighted considerable inconsistencies in the mapping process. These inconsistencies can be attributed to various factors, including variations in the expertise of surveyors and the intricacies of the indicators employed to characterize slum conditions. Consequently, reliable data is lacking, which poses a significant barrier to effective monitoring of slum upgrading programs. Remote sensing (RS)-based approaches, particularly those employing deep learning (DL) techniques, have emerged as a highly effective and accurate method for identifying slum areas. However, the reliance on RS alone is likely to encounter challenges in complex urban environments. A substantial body of research has previously identified the merits of integrating land surface data with RS. Therefore, this study seeks to combine remote sensing imagery (RSI) with street view imagery (SVI) for the purpose of slum mapping and compare its accuracy with a field survey conducted in 2024. The city of Bandung is a pertinent case study, as it is facing a considerable increase in population density. These slums collectively encompass approximately one-tenth of Bandung City’s population as of 2020. The present investigation evaluates the mapping results obtained from four distinct deep learning (DL) networks: The first category comprises FCN, which utilizes RSI exclusively, and FCN-DK, which also employs RSI as its sole input. The second category consists of two networks that integrate RSI and SVI, namely FCN and FCN-DK. The findings indicate that the integration of RSI and SVI enhances the precision of slum mapping in Bandung City, particularly when employing the FCN-DK network, achieving an accuracy of 86.25%. The results of the mapping process employing a combination of the FCN-DK network, which utilizes the RSI and SVI, indicate the presence of 2294 light slum points and 29 medium slum points. It should be noted that the outcomes are contingent upon the methodological approach employed, the accessibility of the dataset, and the training data that mirrors the distribution of slums in 2020 and the specific degree of its integration within the FCN network. The FCN-DK model, which integrates RSI and SVI, demonstrates enhanced performance in comparison to the other models examined in this study.