Mapping with fAIr

Assessing the performance of AI-assisted mapping of building footprints for OSM

Anna Zanchetta, Omran Najjar, Kshitij Sharma

The Alan Turing Institute

HOT Humanitarian OpenStreetMap Team

09/08/2024

Introduction

Omran (Germany)

AI Product owner

Kshitij (Nepal)

Backend developer

fAIr

https://fair-dev.hotosm.org/

RGB from OAM

Open Aerial Map

Labels from OSM

Preprocessing through fAIr website

ML engine

https://rampml.global/

2020 paper by Baheti et al.

Methods

Reason for research

How accurate is fAIr in detecting buildings

in different conditions?

Metrics

Source: https://metrics-reloaded.dkfz.de/metric-library

Metrics

Source: https://metrics-reloaded.dkfz.de/metric-library

Dataset


Urban regions	25
Countries	21
Zoom levels	19, 20, 21
N. images	8400 (~350 per region)
Images size	256x256
Resolution	cm

Locations

Degree of urbanity

Rural

Desa Kulaba

[Indonesia]

Peri-urban

Ggaba

[Uganda]

Urban

Bogota

[Colombia]

Refugee camp

Kakuma

[Kenya]

Roof cover type

Shingles

Silvania

[Brasil]

Metal

Ngaoundere

[Cameroon]

Cement

Melbourne

[Australia]

Mixed

Kutupalong

[Bangladesh]

Urban density

Dense

Montevideo

[Uruguay]

Sparse

Gornja Rijeka

[Croatia]

Grid

Quincy

[USA]

Results

Training


Urban regions	all (25)
N. of epochs	20
Batch sizes	4 (2, 4, 8, 16)
Accuracy metrics	5 Categorical accuracy, Precision, Recall, F1 Score, IoU

~~Urbanity~~

~~Roof type~~

Density

~~Batch size~~

Conclusions

4.8%

4.2%

Conclusions

4.8%

7.7%

ooo

ooo

IoU

4.2%

7.8%

Banyuwangi

[Indonesia]

Pallaby, Dhaka

[Bangladesh]

Future

Models

Features

THANK YOU

Code https://github.com/ciupava/fAIr-utilities
fAIr website https://fair.hotosm.org/
Link to this presentation