Transformer-based semantic segmentation for informal settlement mapping using Sentinel-2 MSI imagery

Abstract

Informal settlements house over one billion people globally, yet they remain chronically undermapped in official datasets, undermining progress towards Sustainable Development Goal 11. While deep learning has advanced automated mapping, convolutional neural networks (CNN) are constrained by local receptive fields that cannot adequately capture the long-range spatial dependencies required to distinguish fragmented settlement clusters from spectrally similar urban backgrounds. This study introduces and validates SegFormer, a Transformer-based semantic segmentation framework, for operational informal settlement mapping using freely available 10m Sentinel-2 MSI imagery. Using a 32-channel feature stack, combining spectral, built-up, textural and morphological features, SegFormer outperforms U-Net and DeepLabV3+ (Dice: 0.9934; mIoU: 0.9914), with only 3.6 million parameters, less than half the size of benchmarked CNNs. The framework reduces systematic boundary underestimation and spectral confusion with non-settlement surfaces, addressing persistent limitations of both traditional machine learning and CNN-based approaches. Thus, this scalable approach has the potential to support monitoring of informal settlement growth, flood risk exposure, and targetted infrastructure upgrading across data-scarce urban environments in the Global South.