TY - JOUR
T1 - Intra-urban land use maps for a global sample of cities from Sentinel-2 satellite imagery and computer vision
AU - Guzder-Williams, Brookie
AU - Mackres, Eric
AU - Angel, Shlomo
AU - Blei, Alejandro M.
AU - Lamson-Hall, Patrick
N1 - Funding Information:
We would like to thank Google's generous support through their Geo for Good Cloud Credits program. The 2016 Atlas of Urban Expansion was made possible by financial support from UN-Habitat, the Lincoln Institute of Land Policy, and the Marron Institute at New York University. WRI received funding from Patrick J. McGovern Foundation and the Bezos Earth Fund for this research. Additionally, we want to acknowledge the National Geographic Society for funding early iterations of this work.
Funding Information:
We would like to thank Google’s generous support through their Geo for Good Cloud Credits program. The 2016 Atlas of Urban Expansion was made possible by financial support from UN-Habitat , the Lincoln Institute of Land Policy , and the Marron Institute at New York University . WRI received funding from Patrick J. McGovern Foundation and the Bezos Earth Fund for this research. Additionally, we want to acknowledge the National Geographic Society for funding early iterations of this work.
Publisher Copyright:
© 2022
PY - 2023/3
Y1 - 2023/3
N2 - Intra-urban land use maps provide important information to urban planners and policymakers, but these maps are costly, time consuming to generate and are often unavailable in developing countries where most urban growth is now occurring. This paper reports on machine learning methods to automate the production of land use maps from cloud-free mosaics of Sentinel-2 imagery. We have trained a novel neural network architecture to produced 5 meter resolution land use maps for a global stratified sample of 200 cities. The sample includes all world regions, 78 countries, and a range of population sizes. The model architecture is roughly 1 to 2 orders of magnitude smaller than similar architectures such as UNet (Ronneberger, Fischer, & Brox, 2015) and DeeplabV3+ (Chen, Zhu, Papandreou, Schroff, & Adam, 2018), significantly lowering the cost and computational requirements to produce maps. We are in the process of generating land use maps for all 4,000 + cities and metropolitan areas in the world with populations exceeding 100,000. The resulting product will be the first, regularly updated, freely available, global intra-urban land use maps at 5 meter resolution. We present a 4-tier land use taxonomy which at its root distinguishes open-space from built-up area. At the second tier, it subdivides the built-up category into nonresidential and residential areas. The third tier distinguishes formal from informal residential land use, and the fourth tier further subdivides formal and informal residential land uses into more detailed categories. Accuracy scores at tier-1 and tier-2 were 86% and 79% respectively. Tiers 3 and 4 had an accuracy scores of 75% and 71% respectively. Additionally, we train a roads-only model and compare its output to the Atlas of Urban Expansion's Arterial Roads dataset and Open Street Map. As an example use case, we train an Informal Settlement Classifier, correctly classifying 87% of the settlements.
AB - Intra-urban land use maps provide important information to urban planners and policymakers, but these maps are costly, time consuming to generate and are often unavailable in developing countries where most urban growth is now occurring. This paper reports on machine learning methods to automate the production of land use maps from cloud-free mosaics of Sentinel-2 imagery. We have trained a novel neural network architecture to produced 5 meter resolution land use maps for a global stratified sample of 200 cities. The sample includes all world regions, 78 countries, and a range of population sizes. The model architecture is roughly 1 to 2 orders of magnitude smaller than similar architectures such as UNet (Ronneberger, Fischer, & Brox, 2015) and DeeplabV3+ (Chen, Zhu, Papandreou, Schroff, & Adam, 2018), significantly lowering the cost and computational requirements to produce maps. We are in the process of generating land use maps for all 4,000 + cities and metropolitan areas in the world with populations exceeding 100,000. The resulting product will be the first, regularly updated, freely available, global intra-urban land use maps at 5 meter resolution. We present a 4-tier land use taxonomy which at its root distinguishes open-space from built-up area. At the second tier, it subdivides the built-up category into nonresidential and residential areas. The third tier distinguishes formal from informal residential land use, and the fourth tier further subdivides formal and informal residential land uses into more detailed categories. Accuracy scores at tier-1 and tier-2 were 86% and 79% respectively. Tiers 3 and 4 had an accuracy scores of 75% and 71% respectively. Additionally, we train a roads-only model and compare its output to the Atlas of Urban Expansion's Arterial Roads dataset and Open Street Map. As an example use case, we train an Informal Settlement Classifier, correctly classifying 87% of the settlements.
KW - Computer vision
KW - Google Earth Engine
KW - Informal settlements
KW - Land use land cover
KW - Neural networks
KW - Sentinel-2
KW - Supervised classification
KW - Urban land use maps
UR - http://www.scopus.com/inward/record.url?scp=85143682108&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85143682108&partnerID=8YFLogxK
U2 - 10.1016/j.compenvurbsys.2022.101917
DO - 10.1016/j.compenvurbsys.2022.101917
M3 - Article
AN - SCOPUS:85143682108
SN - 0198-9715
VL - 100
JO - Computers, Environment and Urban Systems
JF - Computers, Environment and Urban Systems
M1 - 101917
ER -