Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Automated National Urban Map Extraction

Hasan Nasrallah, Abed Ellatif Samhat, Cristiano Nattero, Ali J. Ghandour

TL;DR

This work tackles the lack of national rooftop maps in developing regions by introducing an automated pipeline that performs multi-class building instance segmentation on sub-meter satellite imagery. The approach uses a UNet-like architecture with multi-channel outputs for Building interiors, Borders, and Spacing, coupled with a customized loss to improve boundary delineation and instance separation. A Lebanon-based case study trains on a Tyre city dataset, applying 5-fold cross validation, data augmentation including CutMix, and watershed post-processing to produce a first comprehensive Lebanese urban map of ~1 million buildings with an accuracy around 84%. The results show that border-aware, two-class segmentation with watershed post-processing significantly improves instance separation over single-class methods, and comparisons to OpenStreetMap reveal substantial under-labeling in the baseline maps, underscoring the method’s practical impact for national-scale urban planning and solar rooftop potential estimation.

Abstract

Developing countries usually lack the proper governance means to generate and regularly update a national rooftop map. Using traditional photogrammetry and surveying methods to produce a building map at the federal level is costly and time consuming. Using earth observation and deep learning methods, we can bridge this gap and propose an automated pipeline to fetch such national urban maps. This paper aims to exploit the power of fully convolutional neural networks for multi-class buildings' instance segmentation to leverage high object-wise accuracy results. Buildings' instance segmentation from sub-meter high-resolution satellite images can be achieved with relatively high pixel-wise metric scores. We detail all engineering steps to replicate this work and ensure highly accurate results in dense and slum areas witnessed in regions that lack proper urban planning in the Global South. We applied a case study of the proposed pipeline to Lebanon and successfully produced the first comprehensive national building footprint map with approximately 1 Million units with an 84% accuracy. The proposed architecture relies on advanced augmentation techniques to overcome dataset scarcity, which is often the case in developing countries.

Automated National Urban Map Extraction

TL;DR

This work tackles the lack of national rooftop maps in developing regions by introducing an automated pipeline that performs multi-class building instance segmentation on sub-meter satellite imagery. The approach uses a UNet-like architecture with multi-channel outputs for Building interiors, Borders, and Spacing, coupled with a customized loss to improve boundary delineation and instance separation. A Lebanon-based case study trains on a Tyre city dataset, applying 5-fold cross validation, data augmentation including CutMix, and watershed post-processing to produce a first comprehensive Lebanese urban map of ~1 million buildings with an accuracy around 84%. The results show that border-aware, two-class segmentation with watershed post-processing significantly improves instance separation over single-class methods, and comparisons to OpenStreetMap reveal substantial under-labeling in the baseline maps, underscoring the method’s practical impact for national-scale urban planning and solar rooftop potential estimation.

Abstract

Developing countries usually lack the proper governance means to generate and regularly update a national rooftop map. Using traditional photogrammetry and surveying methods to produce a building map at the federal level is costly and time consuming. Using earth observation and deep learning methods, we can bridge this gap and propose an automated pipeline to fetch such national urban maps. This paper aims to exploit the power of fully convolutional neural networks for multi-class buildings' instance segmentation to leverage high object-wise accuracy results. Buildings' instance segmentation from sub-meter high-resolution satellite images can be achieved with relatively high pixel-wise metric scores. We detail all engineering steps to replicate this work and ensure highly accurate results in dense and slum areas witnessed in regions that lack proper urban planning in the Global South. We applied a case study of the proposed pipeline to Lebanon and successfully produced the first comprehensive national building footprint map with approximately 1 Million units with an 84% accuracy. The proposed architecture relies on advanced augmentation techniques to overcome dataset scarcity, which is often the case in developing countries.
Paper Structure (32 sections, 6 equations, 10 figures, 6 tables, 1 algorithm)

This paper contains 32 sections, 6 equations, 10 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Study Area
  4. Region of Study
  5. Dataset creation
  6. K-fold cross validation
  7. Border Creation
  8. Spacing Creation
  9. Methodology
  10. Model Architecture
  11. Encoder Choice
  12. Input and Output Channels
  13. Building Class
  14. Border Class
  15. Spacing Class
  16. ...and 17 more sections

Figures (10)

  • Figure 1: Number of 1024px Tiles per buildings count.
  • Figure 2: Encoder-Decoder UNet Architecture with Skip Connections
  • Figure 3: Post Processing Method
  • Figure 4: IoU's and Fscore's on Validation Dataset vs Batch Size for each EfficientNet encoder and AMP usage
  • Figure 5: CutMix Image and Mask Creation: A random bounding box is cropped from first image and pasted in the second Image.
  • ...and 5 more figures