Mapping with fAIr

Validation accuracy after training in building footprints segmentation for OpenStreetMap

Anna Zanchetta, Omran Najjar, Kshitij Sharma

The Alan Turing Institute

HOT Humanitarian OpenStreetMap Team

2024-06-25

Introduction

HOT

fAIr

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

ML engine

https://rampml.global/

ML engine

2020 paper by Baheti et al.

Reason for research

How accurate is fAIr in detecting buildings

in different conditions?

Methods

Data

RGB from OAM








Open Aerial Map

Labels from OSM








Preprocessing through fAIr website

Dataset

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

Locations

Categories

Rural

Desa Kulaba

Peri-urban

Ggaba

Urban

Bogota

Refugee camp

Kakuma

Categories

Shingles

Silvania

Metal

Ngaoundere

Cement

Melbourne

Mixed

Kutupalong

Categories

Dense

Montevideo dense

Sparse

Gornja Rijeka

Grid

Quincy

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

Results

Urbanity

Roof type

Density

Batch size

Conclusions

Conclusions

4.8%

4.2%

Conclusions

4.8%

7.7%

ooo

ooo

IoU

4.2%

7.8%

Banyuwangi

Pallaby Dhaka

Future

Models

Features

THANK YOU

GitHub repo https://github.com/ciupava/fAIr-utilities
fAIr website https://fair-dev.hotosm.org/
This presentation