[Paper Review] Mapping Informal Settlements in Developing Countries with Multi-resolution, Multi-spectral Data

Patrick Helber, Bradley Gram-Hansen+5 more|arXiv (Cornell University)|Nov 30, 2018

Remote-Sensing Image Classification6 references4 citations

TL;DR

This paper proposes two methods for mapping informal settlements in developing countries using multi-resolution, multi-spectral satellite data: a computationally efficient Canonical Correlation Forest (CCF) classifier trained on freely available low-resolution (10m) Sentinel-2 imagery, and a deep learning semantic segmentation model using costly very-high-resolution (30cm) imagery. To our knowledge, it is the first study to successfully map informal settlements using only low-resolution satellite data, achieving strong generalization across diverse global regions despite noisy and incomplete annotations.

ABSTRACT

Detecting and mapping informal settlements encompasses several of the United Nations sustainable development goals. This is because informal settlements are home to the most socially and economically vulnerable people on the planet. Thus, understanding where these settlements are is of paramount importance to both government and non-government organizations (NGOs), such as the United Nations Children's Fund (UNICEF), who can use this information to deliver effective social and economic aid. We propose two effective methods for detecting and mapping the locations of informal settlements. One uses only low-resolution (LR), freely available, Sentinel-2 multispectral satellite imagery with noisy annotations, whilst the other is a deep learning approach that uses only costly very-high-resolution (VHR) satellite imagery. To our knowledge, we are the first to map informal settlements successfully with low-resolution satellite imagery. We extensively evaluate and compare the proposed methods. Please find additional material at https://frontierdevelopmentlab.github.io/informal-settlements/.

Motivation & Objective

• To address the critical need for accurate, scalable mapping of informal settlements in developing countries to support UN Sustainable Development Goals.
• To develop a method for detecting informal settlements using only freely available, low-resolution (10m) multi-spectral satellite imagery.
• To evaluate the performance and generalization capability of a low-cost, computationally efficient method (CCF) versus a high-accuracy, high-cost deep learning approach (CNN) on very-high-resolution imagery.
• To demonstrate that spectral signatures alone from low-resolution data can enable effective detection of informal settlements, even in complex urban environments.

Proposed method

• A pixel-wise classification model using Canonical Correlation Forests (CCFs) is trained on 10-band multi-spectral data from Sentinel-2 at 10m resolution to classify each pixel as either informal settlement or environment.
• The CCF model is trained with 80% of annotated data per class, using n_trees = 10, and normalized input features (zero mean, unit variance) to improve generalization.
• For high-precision mapping, a state-of-the-art semantic segmentation convolutional neural network (CNN) is trained on very-high-resolution (30cm) DigitalGlobe imagery to capture fine-grained structural features such as roof size and settlement density.
• Model evaluation uses pixel accuracy and mean Intersection over Union (mIoU), with separate training and test sets constructed from randomly sampled points per class.
• Generalization is tested by training models on one region (e.g., Northern Nairobi or Medellin) and evaluating on unseen regions, demonstrating transfer learning potential.
• The study leverages data from six countries (Kenya, South Africa, Nigeria, Sudan, Colombia, India), with ground truth annotations from UNICEF and other partners.

Experimental results

Research questions

• RQ1 Can informal settlements be effectively detected using only low-resolution (10m) multi-spectral satellite imagery without relying on high-resolution data?
• RQ2 How does the performance of a computationally efficient CCF model trained on low-resolution data compare to a deep learning CNN model trained on very-high-resolution imagery?
• RQ3 To what extent can a model trained on one region generalize to detect informal settlements in geographically and climatically distinct, unseen regions?
• RQ4 What is the impact of noisy and incomplete annotations on model performance in low-data regimes for informal settlement detection?
• RQ5 Can spectral signatures alone from low-resolution data capture the essential characteristics of informal settlements, even when structural details are indistinguishable?

Key findings

• The CCF model trained on low-resolution (10m) Sentinel-2 data achieved a pixel accuracy of 69.4% and mIoU of 62.0% on Northern Nairobi, outperforming the CNN on Kibera despite lower resolution.
• On Kibera, the CCF model achieved 69.0% pixel accuracy and 73.3% mIoU, while the CNN achieved 78.2% and 65.5%, respectively, showing that spectral data alone can capture settlement structure effectively.
• In Al Geneina, Sudan, where spectral signatures are similar to surroundings, the CNN achieved 89.2% accuracy and 76.3% mIoU, demonstrating the advantage of high-resolution features in complex cases.
• The CCF model trained on Northern Nairobi generalized to other regions with moderate success: 69.4% accuracy and 62.0% mIoU on Kibera, and 55.0% accuracy and 54.4% mIoU on Medellin, indicating transfer learning potential.
• In regions with incomplete ground truth (e.g., South Africa, Nigeria, India), the CCF model maintained moderate performance (e.g., 97.0% accuracy in Mumbai), suggesting robustness to data sparsity.
• The study demonstrates, for the first time, that informal settlements can be reliably detected using only freely available, low-resolution multi-spectral satellite data, opening new pathways for scalable, low-cost urban monitoring.