Spatial and deep learning analyses of urban recovery from the impacts of COVID-19

Abstract

This study investigates urban recovery from the COVID-19 pandemic by focusing on three main types of working, commercial, and night-life activities and associating them with land use and inherent socio-economic patterns as well as points of interests (POIs). Massive multi-source and multi-scale data include mobile phone signaling data (500 m × 500 m), aerial images (0.49 m × 0.49 m), night light satellite data (500 m × 500 m), land use data (street-block), and POIs data. Methods of convolutional neural network, guided gradient-weighted class activation mapping, bivariate local indicator of spatial association, Elbow and K-means are jointly applied. It is found that the recovery in central areas was slower than in suburbs, especially in terms of working and night-life activities, showing a donut-shaped spatial pattern. Residential areas with mixed land uses seem more resilient to the pandemic shock. More than 60% of open spaces are highly associated with recovery in areas with high-level pre-pandemic social-economic activities. POIs of sports and recreation are crucial to the recovery in all areas, while POIs of transportation and science/culture are also important to the recovery in many areas. Policy implications are discussed from perspectives of open spaces, public facilities, neighborhood units, spatial structures, and anchoring roles of POIs.