Decoding (urban) form and function using spatially explicit deep learning

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11310%2F24%3A10483562" target="_blank" >RIV/00216208:11310/24:10483562 - isvavai.cz</a>

Výsledek na webu

<a href="https://verso.is.cuni.cz/pub/verso.fpl?fname=obd_publikace_handle&handle=6T_LVou94i" target="_blank" >https://verso.is.cuni.cz/pub/verso.fpl?fname=obd_publikace_handle&handle=6T_LVou94i</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1016/j.compenvurbsys.2024.102147" target="_blank" >10.1016/j.compenvurbsys.2024.102147</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Decoding (urban) form and function using spatially explicit deep learning

Popis výsledku v původním jazyce

This paper explores how can geographical dimension be incorporated into deep learning designed to understand the composition of urban landscapes based on Sentinel 2 satellite imagery. Compared to standard computer vision, satellite imagery is unique as images sampled from the data form a continuous array, rather than being fully independent. We argue that the spatial configuration of the images is as important as the content of each image when attempting to capture a pattern that reflects the structure of the urban environment. We propose a series of approaches explicitly incorporating spatial dimension in the predictive pipeline based on the EfficientNetB4 convolutional neural network (CNN) and experimentally test their effect on model performance. The experiments in this study cover the scale of the sampled area, the effect of spatial augmentation, and the role of modelling (logit ensemble and histogram-based gradient-boosted classifiers) with and without the spatial context on the outputs of the neural network-generated vector of probabilities while trying to predict spatial signatures, a classification of primarily urban landscape based on form and function. The results suggest that certain ways of embedding spatial information, especially in the modelling step, consistently significantly improve the prediction accuracy and shall be considered on top of standard CNNs.

Název v anglickém jazyce

Decoding (urban) form and function using spatially explicit deep learning

Popis výsledku anglicky

This paper explores how can geographical dimension be incorporated into deep learning designed to understand the composition of urban landscapes based on Sentinel 2 satellite imagery. Compared to standard computer vision, satellite imagery is unique as images sampled from the data form a continuous array, rather than being fully independent. We argue that the spatial configuration of the images is as important as the content of each image when attempting to capture a pattern that reflects the structure of the urban environment. We propose a series of approaches explicitly incorporating spatial dimension in the predictive pipeline based on the EfficientNetB4 convolutional neural network (CNN) and experimentally test their effect on model performance. The experiments in this study cover the scale of the sampled area, the effect of spatial augmentation, and the role of modelling (logit ensemble and histogram-based gradient-boosted classifiers) with and without the spatial context on the outputs of the neural network-generated vector of probabilities while trying to predict spatial signatures, a classification of primarily urban landscape based on form and function. The results suggest that certain ways of embedding spatial information, especially in the modelling step, consistently significantly improve the prediction accuracy and shall be considered on top of standard CNNs.

Druh

J_imp - Článek v periodiku v databázi Web of Science

CEP obor

—

OECD FORD obor

50701 - Cultural and economic geography

Projekt

—

Návaznosti

I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název periodika

Computers, Environment and Urban Systems

ISSN

0198-9715

e-ISSN

1873-7587

Svazek periodika

112

Číslo periodika v rámci svazku

September

Stát vydavatele periodika

GB - Spojené království Velké Británie a Severního Irska

Počet stran výsledku

23

Strana od-do

102147

Kód UT WoS článku

001271704800001

EID výsledku v databázi Scopus

2-s2.0-85198291248