Evaluating the social-economic recovery impacts of the built environment post- pandemic: A case study of COVID-19

Abstract

The post-pandemic era presents dual challenges for urban recovery: overcoming its impacts and shaping a new socio-economic pattern. Promoting urban recovery through non-pharmaceutical intervention is crucial. This study examines the impacts of built environment on urban socio-economic recovery in Metropolitan Nagoya in the post-pandemic period. Here we find that the geographical weighted random forest (GWRF) model has a better performance on urban recovery in central districts (kappa value is 67 ± 18%) compared to the peripheral areas in Nagoya, Japan. The balance of jobs and housing pre-pandemic emerges as crucial factor in urban recovery, especially in the low recovery level area. Moreover, accessibility to secondary roads notably promotes urban recovery in central area, while this factor also introduces the risk of lower recovery levels. By integrating nonlinear and spatial methodology, this study offers policymakers crucial insights into how different built environment can impact urban socio-economic recovery and environmental sustainability as the pandemic becoming frequently.

Keywords

Urban recovery post-pandemic built environment geographical weighted random forest spatial heterogeneity

Introduction

The outbreak of the novel coronavirus (COVID-19) unexpectedly disrupted global economic and social progress, especially in densely populated major cities (Wang et al., 2022). To address the urgent needs of epidemic control, many cities implemented strict measures such as lockdowns and work stoppages, which played a crucial role in containing the rapid and widespread spread of the virus. Consequently, these epidemic control policies undoubtedly led to a significant reduction in social and economic activities, culminating in urban decline. As the WTO declared that COVID-19 no longer constitutes a public health emergency of international concern, governments have initiated economic and social recovery plans for the post-pandemic period. Given that cities are typically hotspots for COVID-19 infections due to their high concentration of population and economic activities, urban recovery has become a key element of society’s overall recovery. The post-COVID-19 recovery of cities should be assessed by studying the impact of various factors related to COVID-19 transmission and closely linked to the urban built environment.

Existing studies have mostly focused on the impact of the built environment on the spread of pandemic, but the built environment is equally crucial for urban recovery in the post-pandemic pandemic. The enduring interaction between the built environment and infectious diseases has been evident since the beginning of urbanization, making the urban built environment a key determinant of urban recovery, especially after the devastation caused by pandemics (Frumkin, 2021). The urban built environment influences the city’s recovery from COVID-19 to varying degrees. These built environment factors include the distribution of parks, the distribution of public transportation facilities, the number of campuses, the density of residential land use, land use, open space, and sports and recreational facilities (Keenan, 2020). Existing studies have largely focused on the relationship between spatial characteristics and post-pandemic urban recovery. For example, Ma et al. (2023) find that suburban regions with lower population densities have achieved a higher degree of recovery when compared to central areas. There is a correlation between nightlife and commercial activities and a higher level of recovery (Ma et al., 2021). Areas with many non-residential buildings recover more slowly. For areas with high levels of socio-economic activity, more large parks and open spaces facilitate urban recovery (Dong et al., 2022), while transportation services have unequal contributions to urban recovery (Li et al., 2023). Expanding research on the built environment’s impact on post-epidemic urban recovery is essential, which can provide targeted non-pharmaceutical interventions during both the epidemic transmission and recovery phases, thereby enhancing urban resilience more precisely.

As for factors used to explain the urban recovery during the COVID-19 pandemic, government support, land use, socio-economic indicators, and urbanization levels are often used (Chu et al., 2021; Dueñas et al., 2021; Xu et al., 2021). However, urban environments are complex systems that benefit from a broader range of variables. For example, research highlighting the importance of green space accessibility in London after the COVID-19 pandemic pointed out its role in alleviating psychological distress (Lee and Lee, 2021). Additionally, studies in Northeast China have shown the significance of industrial ratios and governance efficacy in regional economic resurgence (Hu et al., 2022). These findings emphasize the need for a more comprehensive perspective that includes built environment factors alongside economic and demographic components.

Advances in machine learning offer robust alternatives for uncovering latent patterns in complex datasets, which are widely used in urban studies. Techniques like random forests and support vector machines (Nandam and Patel, 2021; Niu et al., 2020) have proven effective in various contexts. Moreover, deep learning has gained traction for tasks like urban growth modeling and land use classification, utilizing methods such as artificial neural networks and convolutional neural networks (Maithani, 2009; Vali et al., 2020). Despite their predictive capabilities, many machine learning models still neglect geo-related variables, particularly spatial heterogeneity, and dependence. Spatial dependence refers to the interrelations between observations within geographic spaces, suggesting that proximity usually implies greater correlation (Smith and LeSage, 2004). Spatial heterogeneity captures variations at different locations, often stemming from natural processes and human activities (Grimm et al., 2000). These overlookings can lead to diminished accuracy in models, underscoring the need for their integration in quantitative urban analysis.

In this study, we used a combination of mobility data and multi-source spatial data to conduct a comprehensive analysis investigating the impact of the built environment on urban socio-economic recovery post-pandemic. To address concerns of spatial heterogeneity, we employed the GWRF approach and investigated how these effects vary across different thresholds. By doing so, our study aims to provide valuable insights into the fine-grained intervention of the built environment that promotes urban recovery from pandemics as a strategy for improving urban resilience.

Data and methods

Study area

Our research focuses on the whole Nagoya City in Japan, as shown in Figure 1. Nagoya is the fourth most populous city proper, with a population of 2.3 million in 2020, and the principal city of the Chukyo metropolitan area, which is the third most populous metropolitan area in Japan. We selected the period after the lifting of the national state of emergency declaration and compared it with the same period in 2019 as the baseline. Specifically, the Nagoya government declared a state of emergency for the first time to prevent the further spread of COVID-19 from April 10, 2020. This declaration led to zero daily infections in Nagoya on May 15, 2020, lasting for 23 days. In this study, the period from May 26 (Tuesday) to June 1 (Monday), 2020, was selected to represent a recovery phase, marking the first week after the lifting of the emergency declaration with no new reported infections. For comparison, the period from May 28 (Tuesday) to June 3 (Monday), 2019, was chosen to represent the pre-COVID-19 baseline.

Figure 1.

Distribution of Nagoya City districts.

Research data

Mobile phone signaling data: Mobile phone signaling data can record spatio-temporal human activity trajectories. The data were obtained from NTT DOCOMO company, which is the largest mobile phone company in Japan. This company had 82.63 million mobile phone users by March 2021, accounting for 43.8% of the mobile phone market in Japan. Original point-to-point data were aggregated into 500 m × 500 m meshes for release to the public, where the total number of population in each mesh is calculated at the hour level. There are a total of 1383 meshes in Nagoya, covering an area of 345.75 km². To mitigate short-term fluctuations in mobile phone signaling data, we aggregated mobile phone users’ data from 09:00 to 17:00, Monday through Friday, to represent human activity patterns.

Land use type data: Land use data were obtained from the City Planning Basic Survey of Land Use in Nagoya in 2017 (https://www.geospatial.jp/ckan/dataset/nagoya-kiso), covering 4086 street-blocks over the whole city. This data set includes percentages of 13 types of land use at the street-block level and was recalculated into mesh level according to its area ratio. It includes ratio of field land (R_field), ratio of forest land (R_forest), ratio of river land (R_river), ratio of other natural land (R_ONL), ratio of residential land (R_RES), ratio of commercial land (R_COM), ratio of industrial land (R_LND), ratio of public facility land (R_PFL), ratio of road land (R_road), ratio of transportation facility land (R_TFL), ratio of public open space(R_POS), ratio of other public facility land (R_OPFL), and ratio of other open space land (R_OOSL).

Point of Interest (POIs) data: There are 39,761 POIs from 263 categories in Nagoya, as shown in data from the 2018 ESRI Japan Corporation. POIs data were converted into 16 large categories: tourism attraction, transportation service, governmental & social group, culture & education, medical service, science/culture service, sports & recreation, food & beverages, shopping, auto sale, auto service, finance service, daily life service, accommodation service, service apartment, and enterprise. Categories in POI that are closely related to the study were singled out and their percentage of the total POI in each 500 m × 500 m grid was calculated. These POI categories include proportion of commercial POIs (P_COM), cultural POIs (P_CUL), educational POIs (P_EDU), entertainment POIs (P_ENT), Greenland POIs (P_GRE), JR station POIs (P_JR), public service facility POIs (P_PSF), and other type of POIs (P_OT).

Road accessibility data: The road data were downloaded from Open Street Map (OSM), including all road information in Nagoya City. These data were imported into ArcGIS Pro for processing. The primary, secondary, and tertiary levels of roads were merged, with intersecting lines broken and topology corrected. The travel speeds were set at 60 km/h for primary and secondary roads and 50 km/h for tertiary roads to build an origin-destination (OD) cost matrix. For every point in the road network, the sum of the travel times for all paths starting from that point was calculated. Then this sum of time was divided by the value of the number of paths starting from that point. The result of this division was used as an indicator of the node’s accessibility, thereby obtaining the accessibility value of every node in the network. With inverse distance weighting interpolation, this accessibility value was extended to every point in space, ultimately obtaining the road accessibility data. This data contains primary road accessibility (access_prim), secondary road accessibility (access_seco), and tertiary road accessibility (access_tert).

Land price level: The land price data were obtained from the Ministry of Land, Infrastructure, Transport and Tourism of Japan (https://nlftp.mlit.go.jp/ksj/gml/datalist/KsjTmplt-L01-v3_1.html), which included the 2020 land price levels in Nagoya City and the rate of change compared to the previous year. The land price data, of point type, comprised 1902 data points from the entire city of Nagoya. After being imported into ArcGIS Pro, the data points were connected to the covering spatial grid using spatial join tool, and the average value of land prices within each grid was calculated as the average rent.

Research methods

Measurement of urban recovery: The urban recovery from COVID-19 was measured with respect to the daytime activities at the mesh level, where the main time slots of each day were selected: i.e., 09:00–17:00 for human activities from 26 May to 1 June in 2020 comparing 28 May and 3 June in 2019. To verify the stability of the weekly data during the study period, we calculated the Pearson correlation coefficient between human activities over the 90 days following the reopening and those during the study period, yielding a value of 0.98, which was significant at the 0.05 level. The change of the number of populations comparing recovery period and baseline was been urban recovery condition. The specific calculation formula is as follows:

U R_{i} = \frac{H A_{i, 2020}}{H A_{i, 2019}}

Here UR_i represents the urban recovery condition in the $i_{t h}$ mesh, and $H A_{i, 2020}$ and $H A_{i, 2020}$ refer to the human activity levels during the recovery phase and pre-COVID-19 phase in the $i_{t h}$ mesh, respectively. To facilitate data processing and reveal the relationship between recovery levels and associated factors, $U R_{i}$ is classified into five levels (from 1, lowest, to 5, highest) using the equal quantile method. Given that human activities in residential areas do not accurately reflect socio-economic recovery post-pandemic, meshes primarily composed of residential areas were excluded from this study. Following Liu and Long’s (2016) method, meshes with over 50% residential land use (totaling 1383 meshes) were identified and removed, resulting in a study area comprising 1208 meshes. The geographical distribution of the measured urban recovery level is shown in Figure 3a.

Measurement of job-housing balanced degree: The conventional measure of job-housing balance is to calculate the ratio of employment opportunities to the number of households, but obtaining precise data on employment opportunities can be challenging. Since people typically work in offices during the day and rest at home at night, the ratio of daytime to nighttime mobile signaling can partially reflect the job-housing balanced degree (Dong et al., 2022). A study conducted in the Greater Stockholm metropolitan area demonstrates that the ratio of mobile phone signaling data between daytime and nighttime periods serves as an effective indicator of the job-housing balance degree (Müürisepp et al., 2023; Toger et al., 2020). By dividing the area into relatively conservative 500 m × 500 m grid cells, the study effectively mitigated the interference of daytime and nighttime locational uncertainties of mobile phone users on the research data. (Ogulenko et al., 2022; You and Guan, 2024). Considering that our research focuses on the overall recovery of urban areas, with a particular emphasis on non-residential zones to accurately capture the restoration of commercial and industrial activities, we excluded grids concentrated in residential land-use areas. This removal of predominantly residential grids applies solely to resilience analysis, aiming to avoid the interference effects caused by increased home-based activities during the COVID-19. This study calculated the ratio of nighttime to daytime mobile signaling data (R_ND_19) for a working day in 2019 and the same period (also a working day) in 2020 (R_ND_20), which represents the normal level of job-housing balanced degree and the special level of it during the COVID-19 respectively. The larger the deviation from a value of 1 in the calculated signaling ratio, the lower the degree of job-housing balance. The difference value fails to capture the baseline (2019) job-housing balanced degree, which could be a crucial factor influencing recovery patterns (Ma et al., 2023). Identical difference values may represent distinctly different recovery patterns when associated with different baseline levels. By employing both the ratio and metrics from two independent years, we can capture nonlinear relationships and separately evaluate the impacts of pre-pandemic job-housing balanced degree and its COVID-19-induced changes on urban recovery, thus avoiding oversimplification of complex spatio-temporal relationships.

Geographical weighted random forest (GWRF): GWRF decomposes the Random Forest (RF) model into several geographic sub-models, meaning that for each location $i$ , a local random forest sub-model (labeled $R F_{i}$ ) is computed, with only the observed data within a radius of n around location $i$ involved in training. Once training is completed, GWRF combines the predictions of both the global and local models during the prediction phase, introducing weights ( $w$ ) that allow us to control the influence of the global and local models on the model output to different extents. GWRF model is a local, nonlinear spatial machine learning method that identifies key factors influenced by their surrounding environment (Georganos et al., 2019; Luo et al., 2021). This model has shown superiority over Geographical Weighted Regression (GWR) and other linear spatial models in handling spatial heterogeneity. For instance, a study using the GWR model in the US highlighted region-specific critical factors affecting COVID-19 deaths, such as lack of medical insurance in the South and Midwest, and the absence of recreational sports activities on the East Coast (Grekousis et al., 2022). In this study, a geographical weighted random forest model consisting of 1208 local random forests was trained and its accuracy is validated using existing data. By incorporating spatial variables, we have significantly enhanced the predictive accuracy of urban recovery in our model. The main hyperparameter of the GWRF model is bandwidth, which controls the distance over which the random forest model is influenced by surrounding variables (larger bandwidth allows the random forest model to incorporate data from a larger range of neighboring parcels). In the hyperparameter analysis, we adjusted the bandwidth to determine the optimal model. Since the GWRF model trained in the research is a classification model (with the dependent variable being recovery level, divided into five categories), the kappa coefficient is used to measure the model’s accuracy. The kappa coefficient assesses inter-rater agreement for categorical data, adjusting for chance agreement. It compares observed agreement with expected agreement, ranging from −1 (complete disagreement) to 1 (complete agreement), with 0 indicating agreement due to chance. In this study, model construction and accuracy evaluation were implemented using GWRFC packages in R, which is put forward by Santos et al. (2019). Table 1 shows the variables used in GWRF model.

Table 1.

The variables used in GWRF model.

Variables	The meaning of variable	Basis for data selection	Variables	The meaning of variable	Basis for data selection
access_prim	Accessibility to the primary roads	Forouhar et al. (2024)	R_field	The ratio of field land	Santiago-Iglesias et al. (2024)
access_seco	Accessibility to the secondary roads	Forouhar et al. (2024)	R_forest	The ratio of forest	Santiago-Iglesias et al. (2024)
access_tert	Accessibility to the tertiary roads	Forouhar et al. (2024)	R_IND	The ratio of industrial land	Forouhar et al. (2024)
P_COM	Proportion of commercial POI	Ma et al. (2023)	R_ONL	The ratio of open natural land	Dong et al. (2022)
P_CUL	Proportion of cultural POI	Ma et al. (2023)	R_OOSL	The ratio of other open space land	Dong et al. (2022)
P_EDU	Proportion of educational POI	Ma et al. (2023)	R_OPFL	The ratio of other public facility land	Dong et al. (2022)
P_ENT	Proportion of entertaining POI	Ma et al. (2023)	R_PFL	The ratio of public facility land	Dong et al. (2022)
P_GRL	Proportion of entertaining POI	Ma et al. (2023)	R_POS	The ratio of public open space	Dong et al. (2022)
P_JR	Proportion of JR (Japanese railway) related POI	Ma et al. (2023)	R_RES	The ratio of residential land	Dong et al. (2022)
P_PSF	Proportion of public service facility POI	Ma et al. (2023)	R_river	The ratio of river	Forouhar et al. (2024)
R_ND_19	Ratio of nighttime to daytime signaling data in 2019	Müürisepp et al. (2023); Santiago-Iglesias et al. (2024)	R_road	The ratio of road	Forouhar et al. (2024)
R_ND_20	Ratio of nighttime to daytime signaling data in 2020	Müürisepp et al. (2023); Santiago-Iglesias et al. (2024)	R_TFL	The ratio of traffic facility land	Dong et al. (2022)
rent	The land price	Yang et al. (2020)	dist_subway	Distance to the closest subway station	Dong et al. (2022)
R_COM	The ratio of commercial land	Santiago-Iglesias et al. (2024)	dist_JR	Distance to the closest JR station	Dong et al. (2022)

Local variable importance (LVI) and visualization: Local variable importance is the significance of the target variable within each grid cell. During training, each 500 m × 500 m grid cell is used to train a local random forest model which, like traditional random forests, carries information on the importance of different variables. We calculated the GINI coefficient of each variable in all local sub-models, which were thereby summarized as local variable importance (LVI). The visualization process is performed through ArcGIS Pro, where the LVI is overlaid with Nagoya’s geographic boundaries and the 500 m × 500 m grid cells.

Partial dependence plot (PDP): Each 500 m × 500 m grid cell corresponds to a local random forest model. Assuming there is a trained model that can predict a target variable, and we are interested in a particular feature, PDP can help researchers understand the mechanism of the feature’s impact on the model’s predictions. First, a feature is selected along with a set of values to evaluate. To calculate partial dependence, the values of all other features need to be held constant, changing only the value of the feature of interest. The impact of the feature on the model’s predictions is thereby assessed without interference from other features. In the process mentioned above, a new set of data points is generated, in which only the values of the feature of interest change, while the values of all other features remain the same as in the original data. Then, for each new data point, predictions are made using the trained model, and the results are recorded, which gave us a set of model predictions that correspond to the values of the feature of interest. Since our geographical weighted random forest is a large model comprising 1208 small Random Forest models, to better represent the PDP, we average the results and plot the values at the upper quartile and lower quartile, resulting in the corresponding box plot.

Results

The performance of GWRF

In the GWRF model, we systematically investigated the impact of varying the model’s bandwidth on predictive accuracy. This exploration involved tuning the bandwidth parameter to delineate the optimal size for the range of local sub-model learning. Each GWRF model underwent training with bandwidth values ranging from 0 to 1200 in increments of 50. The resulting box plots of kappa coefficients for each model, presented in Figure 2, revealed insightful patterns.

Figure 2.

Box plot of the variation of kappa coefficient with bandwidth.

Initially, a discernible decreasing trend in the kappa coefficient was observed as the bandwidth increased from 0 to 200. However, beyond a bandwidth of 200, there was a subsequent upswing in the kappa coefficient, reaching a local peak at a bandwidth of 400. As the bandwidth surpassed 400, the kappa coefficient exhibited slight fluctuations with increasing bandwidth, eventually demonstrating a pronounced downward trajectory after surpassing 600. Notably, within the bandwidth range of 1000 to 1500, the model’s kappa coefficients displayed a cliff-like decrease.

In the context of enlarging the bandwidth of the GWRF model, which essentially expands the learning range of the local model, the observed drop in the kappa coefficient beyond a bandwidth of 600 and the cliff-like decrease around the 1000 bandwidth mark suggested that data outside a certain range surrounding the local unit might exert a negative influence on the training of the local sub-model. This implication raises the possibility that units situated beyond the spatial range are relatively weakly connected to the local unit.

Given that an increase in bandwidth resulted in an approximately linear increase in computation (Georganos et al., 2021), and considering that model accuracy did not exhibit a significant enhancement beyond a bandwidth larger than 400, we opted to proceed with the model configured with a bandwidth of 400 for the subsequent discussion. With a suitable bandwidth in place, we proceeded to visualize the spatial distribution of the kappa coefficient for the model, as illustrated in Figure 3.

Figure 3.

Geographical distribution of observed recovery levels (a) and model accuracy (b).

Observing the spatial patterns, it is evident that in the Midori District (kappa value range 38 ± 14%) and Tenpaku District (kappa value range 37 ± 14%), situated in the southeast part of the city, as well as in the Kita District (kappa value range 24 ± 14%) and Nishi District (kappa value range 22 ± 14%), located in the northwest part of the city, the model’s predictions exhibited lower accuracy, ranging from fair to moderate. Conversely, in other areas of the city, the model demonstrated a relatively high level of predictive accuracy. Notably, in the Atsutaku District, positioned at the center of the region, the model’s predictions exhibited the highest consistency with the actual situation.

Visualizing and analyzing feature importance

To present the distribution of feature importance at the global scale, we generated radar plots depicting the mean values of feature importance for each sub-model, categorized by urban recovery levels. In terms of land-use patterns, shown in Figure 4a, it was evident that, in general, the effects of features on different levels of recovery did not exhibit significant differences. For nearly all features, the impact on lower recovery levels was slightly higher than that on higher recovery levels, with the influence reducing as recovery levels increase. Nonetheless, there were substantial variations in the importance attributed to each distinct feature. The plot highlighted the vital role of field land (R_field) in determining recovery levels, whereas nature lands (R_forest, R_ONL, R_river) exhibited comparatively lower significance in the decision-making process for recovery outcomes.

Figure 4.

Radar plot for feature importance: (a) feature related to land use type; (b) other type of features. C1 to C5 refer to the recovery level, from the lowest to the highest.

Regarding features excluding land-use patterns, as illustrated in Figure 4b, the influencing pattern of variables on different recovery levels largely remained consistent. However, the gap between the effects of variables on distinct recovery levels diminished further. Notably, the ratio of nighttime to daytime mobile phone signaling data (R_ND_19, R_ND_20) distinctly stood out, especially the ratio data in 2019, exhibiting nearly the highest impact on every recovery level. Additionally, it was evident that the influence of traffic conditions (access_prim, access_seco, access_tert, dist_subway, dist_JR) should not be underestimated, with the accessibility of secondary roads (access_seco) ranking third after R_ND_19 and R_ND_20.

Subsequently, we mapped the spatial distribution of variable importance for four key features: R_ND_19, R_ND_20, R_field, and access_seco, given their prominence. Focusing on R_ND_19, as depicted in Figure 5a, high values of feature importance were clustered around the northeast, east, and middle parts of the city. Notably, this feature exerted a more significant impact on the highest recovery level, although the importance value exhibited a wider range of variability. The distribution of R_ND_19 displayed positive skewness, with data widely spread along the positive axis. This suggested that the influence from the feature importance was relatively decentralized. However, in districts such as Meitō (importance value range 0.99 ± 0.71) and Moriyama (importance value range 0.74 ± 0.84), where the average feature importance was notably higher, the job-housing balance emerged as a vital factor in promoting urban recovery.

Figure 5.

Spatial distribution of feature importance: (a) R_ND_19; (b) R_ND_20; (c) acces_seco; (d) R_field.

As for R_ND_20, as shown in Figure 5b, the region with high importance values centered around the north, southwest, and south parts of the city, with the highest values concentrated in the south. Notably, the impact of R_ND_20 was more pronounced in predicting lower recovery levels. In contrast to R_ND_19, the distribution of feature importance for R_ND_20 displayed a clear characteristic of negative skewness, indicating that although R_ND_20 had a broader spatial impact on urban recovery than R_ND_19, there were locations, such as the Kita District (importance value range 0.95 ± 0.65), where the feature had a very low impact that should be considered separately. the spatial distribution of R_ND_20 and its effect on different recovery levels markedly differed from R_ND_19. This unexpected divergence raised the possibility of spatial-temporal heterogeneity in the impact of job-housing balance. In districts like Minami (importance value of R_ND_20 range 0.81 ± 0.70, importance value of R_ND_19 range −0.17 ± 0.45) and Minato (importance value of R_ND_20 range 0.71 ± 0.63, importance value of R_ND_19 range −0.41 ± 0.38), the pre-COVID-19 job-housing condition exerted the dominant impact on recovery. However, in districts such as Moriyama (importance value of R_ND_20 range −0.07 ± 0.55, importance value of R_ND_19 range 0.74 ± 0.84), the post-COVID-19 job-housing condition played the decisive role. Despite the overall larger impact of the pre-COVID-19 job-housing condition, the job-housing condition in the post-COVID-19 period had more potential for management. Therefore, greater attention should be directed towards districts such as Moriyama, where the effect of R_ND_19 was relatively marginal, while R_ND_20 took dominance in influencing the recovery level.

In terms of acces_seco, which referred to the accessibility to secondary roads, the region with the highest importance values predominantly concentrated in the central part of the city. Additionally, in the northeastern and eastern parts of the city, the feature exhibited elevated importance values, as shown in Figure 5c. The distribution of feature importance displayed a characteristic of negative skewness, with a small portion of the data widely distributed on the negative axis. This distribution suggested that, at the overall urban level, enhancing accessibility to roads can significantly influence urban recovery, with some exceptions that cannot be neglected. Interestingly, the region where importance value came highest is also the heart of the city area with a large population and high land rent, which showed that the traffic condition should be more highly valued and carefully managed, especially in an area with concentrated population and economic activity.

As for the percentage of field land in land use (R_field, shown in Figure 5d), it can be seen that regions with high importance values were concentrated in the southern and northeastern parts of the city, where field land constituted a relatively small proportion of the land use pattern. Conversely, in the southeastern and western areas of the region, where field land was more prevalent, the impact of this feature was adversely attenuated. This feature exhibited a more pronounced impact on lower recovery levels, indicating that field land held substantial potential to enhance urban recovery, particularly in regions with lower recovery levels.

The nonlinear relationship between the most important built environment and urban recovery

In order to gain a deeper understanding of mechanisms by which features affect recovery levels, we plotted the partial dependencies for the four features with the highest importance values: R_ND_19, R_ND_20, R_field, and access_seco, which illustrated how the predicted probability of each recovery level changed as the feature values varied. According to the Figure 6a, the recovery level exhibited high sensitivity to the R_ND_19 value when it varied around 3. Specifically, the predicted probability of recovery level 2 overwhelmingly dominated when the value was close to but lower than 3. After the value exceeded 3, the predicted probabilities of recovery levels 4 and 5 dramatically increased, accompanied by a decline in the predicted probabilities of levels 1, 2, and 3. Notably, recovery level 5 became dominant after the feature value reached 5. Notably, when the feature value was lower than 1, the recovery level 1 took an absolute dominance, and after the feature value exceeded 2, the dominant level was replaced by level 2. This result suggested that in situations where there was a jobs-housing mismatch, and there were more jobs than available housing in the locality, the locality was much more likely to experience poor recovery. However, when the housing function of the place predominated, with the number of housing units significantly exceeding four times the number of jobs, the jobs-housing mismatch became beneficial to urban recovery.

Figure 6.

Partial dependency plot of the four features with highest importance value: (a) R_ND_19; (b) R_ND_20; (c) Acces_seco; (d) R_field. Darker colors represented higher levels of urban recovery.

The partial dependency of R_ND_20, as shown in Figure 6b, was similar to that of R_ND_19; however, recovery levels appeared to be more sensitive to R_ND_19. When the feature value was less than 2, the recovery level 1 took the dominant status. After the feature value exceeded 2, the predicted probability of level 1 decreased dramatically, while the predicted probabilities of levels 2, 3, 4, and 5 began to rise. Level 2, level 3, and level 4 reached their peak around a feature value of 2.5, 3.5 and 4 respectively, when they briefly dominated before decreasing. The predicted probability of level 5 then took continuous dominance soon after the feature value exceeded 4. The pattern of the partial dependency of R_ND_20 resembled that of R_ND_19: the effect of jobs-housing mismatch had a characteristic of duality; when the jobs predominated, the impact was negative and when the housing was significantly dominant with number of housing units being more than four times the number of jobs, the impact turned to be beneficial, implying that more attention should be paid on the construction of urban residential function.

For acces_seco, the predicted probability of both the highest and lowest recovery levels (level1 and level5) was remained relatively low as the feature value changed. As the accessibility increased and the feature value lowered to a point when the value was less than the threshold of around 4, the gap between the predicted probability of all recovery levels decreased: both the probability of the highest and lowest recovery level (level 5 and level 1) as well as level 2 increased, and the predicted probability of level 3 and level 4 dramatically decreased from a co-dominant position, as shown in Figure 6c. This pattern of partial dependency showed that although the improvement of accessibility to the secondary road could have a significant contribution to the highest recovery level (level 5), there could also be an increasing risk of presenting the lowest recovery level (level 1), and notably the accessibility of the level of recovery with dominant prediction probability decreased from level 3 and level 4 to level 2 and level 3 with the feature value increasing, showed a decreasing trend in the level of recovery, implying that the decision to build roads to improve road accessibility should be taken after careful consideration, as the risks to urban recovery caused by the dense road network should not be ignored.

Regarding the feature of R_field, shown in Figure 6d, the threshold of the feature value is very near to 0. When the feature value was very close to 0, recovery level 2 dominated. After feature value exceeding the threshold, the predicted probability of level 3 and level 5 sharply increased, with the predicted probability of level 1, level 2, and level 4 decreasing dramatically. Recovery level 3 soon took dominant status after the feature value exceeding the threshold. This close-to-zero threshold indicated that the key factor influencing the impact of field land was whether there was any field land present, rather than the quantity being large or small. In districts like Minato, where the importance of field land to urban recovery was highly valued, it would be beneficial to appropriately increase the proportion of field.

Conclusion and discussion

In the face of frequent urban disasters such as pandemics, enhancing urban resilience from the perspective of the built environment is crucial. This study employs a GWRF model and PDP analysis, accounting for spatial heterogeneity and nonlinearity, and integrates multi-source big data on the built environment, socio-economic factors, and human mobility. The research investigates the relationship between the built environment and urban recovery in the post-pandemic context. Our findings indicate that the job-housing balance significantly impacted urban recovery during the post-pandemic period, consistent with findings from metropolitan areas such as Wuhan and California (Islam and Saphores, 2023; Wu et al., 2024). While job-housing balance has been uniformly emphasized in urban studies (Zhang et al., 2022), our results reveal its heightened importance in areas with lower recovery conditions during the post-pandemic era. Secondary road accessibility had a dual impact on urban recovery patterns. While it significantly enhanced recovery performance in urban centers with high accessibility, it paradoxically contributed to minimal recovery in some regions. Additionally, field land played a critical role in promoting urban recovery in areas with higher recovery levels. While prior studies have focused on the quality of green spaces (Chen et al., 2022; Ciesielski et al., 2024; Park et al., 2020), our PDP results further demonstrate that once the ratio of field land exceeds 0.05, its influence on recovery stabilizes. Beyond this threshold, further increases in field land size contribute little additional benefit.

This study broadens the prevailing focus of existing research, which has primarily emphasized medical interventions, to include a perspective centered on built environment factors in the context of non-medical interventions. It addresses a significant gap in urban resilience research, where the role of the built environment as an intervention element has been underexplored. By considering the nonlinear and spatially heterogeneous relationships between the built environment and socio-economic recovery in the post-pandemic era, this study develops targeted policy recommendations. These strategies are tailored to the specific spatial contexts of urban units, balancing the diversity of individual units with the goal of facilitating overall urban recovery, while accounting for intrinsic attributes and surrounding neighborhood spatial relations. Our findings contribute to the theoretical understanding of urban resilience as a non-normative concept (Elmqvist et al., 2019).

This study also has several limitations. First, our research focuses exclusively on Nagoya and does not compare the recovery processes of similar international cities, limiting its ability to provide comprehensive insights into post-pandemic urban recovery strategies. Second, future studies should integrate both the initial pandemic spread and post-pandemic urban recovery to develop systematic strategies for transforming the built environment. Third, while the GWRF model employed in this study offers robust analytical capabilities, it is computationally intensive and difficult to parallelize due to its reliance on Random Forest, which poses challenges for scalability and efficiency.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: We are grateful for the financial support of the National Natural Science Foundation of China (NSFC), Basic Science Center Program, Multiphase Evolution in Hypergravity (No. 51988101).

ORCID iDs

Shuang Ma

Xuanyu Zhou

Wei Cai

Shuangjin Li

Author biographies

Shuang Ma is a Hundred Talents Program Professor at Zhejiang University. Her research focuses on urban resilience, urban planning, and quantitative urban studies.

Xuanyu Zhou is an undergraduate student at Zhejiang University interested in urban sensibility and equity. His research leverages machine learning and big data to address challenges in these areas.

Wei Cai is an undergraduate student at Zhejiang University with an interest in applying artificial intelligence and big data in urban planning and governance.

Mo Chen is a PhD student at Zhejiang University focusing on urban resilience, urban climate, and climate gentrification.

Shuangjin Li is an Assistant Professor at Hiroshima University. Her research focuses on public space quality and environmental health.

References

Chen

Liu

, et al (2022) Change of residents’ attitudes and behaviors toward urban green space pre- and post- COVID-19 pandemic. Land 11(7): Article 7. DOI: 10.3390/land11071051.

Chu

Cheng

Song

(2021) What determines urban resilience against COVID-19: City size or governance capacity? Sustainable Cities and Society 75: 103304.

Ciesielski

Gołos

Stefan

, et al (2024). Unveiling the essential role of green spaces during the COVID-19 pandemic and beyond. Forests 15(2): Article 2. DOI: 10.3390/f15020354.

Dong

Sun

Zhang

(2022) Urban recovery from the COVID-19 pandemic in Beijing, China. The Professional Geographer 74(2): 254–264.

Dueñas

Campi

Olmos

(2021) Changes in mobility and socioeconomic conditions during the COVID-19 outbreak. Humanities and Social Sciences Communications 8: 101.

Elmqvist

Andersson

Frantzeskaki

, et al (2019) Sustainability and resilience for transformation in the urban century. Nature Sustainability 2: 267–273. DOI: 10.1038/s41893-019-0250-1.

Forouhar

Chapple

Allen

, et al (2024) Assessing downtown recovery rates and determinants in North American cities after the COVID-19 pandemic. Urban Studies. Online first. DOI: 10.1177/00420980241270987.

Frumkin

(2021) COVID-19, the built environment, and health. Environmental Health Perspectives 129(7): 75001.

Georganos

Grippa

Niang Gadiaga

, et al (2021) Geographical random forests: A spatial extension of the random forest algorithm to address spatial heterogeneity in remote sensing and population modelling. Geocarto International 36(2): 121–136.

10.

Grekousis

Feng

Marakakis

, et al (2022) Ranking the importance of demographic, socioeconomic, and underlying health factors on US COVID-19 deaths: A geographical random forest approach. Health and Place 74: 102744.

11.

Grimm

Grove

Pickett

STA

, et al (2000) Integrated approaches to long-term studies of urban ecological systems. In: Marzluff

Shulenberger

Endlicher

, et al. (eds) Urban Ecology. Boston, MA: Springer, 123–141.

12.

Dong

(2022) What matters for regional economic resilience amid COVID-19? Evidence from cities in Northeast China. Cities 120: 103440.

13.

Keenan

(2020) COVID, resilience, and the built environment. Environment Systems and Decisions 40(2): 216–221.

14.

Lee

(2021) Public responses to COVID-19 case disclosure and their spatial implications. Journal of Regional Science 62(3): 732–756.

15.

Wei

Zhang

, et al (2023) Applying an interpretable machine learning framework to study mobility inequity in the recovery phase of COVID-19 pandemic. Travel Behaviour and Society 33: 100621.

16.

Liu

Long

(2016) Automated identification and characterization of parcels with OpenStreetMap and points of interest. Environment and Planning B: Planning and Design 43(2): 341–360.

17.

Luo

Yan

McClure

(2021) Distribution of the environmental and socioeconomic risk factors on COVID-19 death rate across continental USA: A spatial nonlinear analysis. Environmental Science and Pollution Research 28(6): 6587–6599.

18.

Zhang

(2021) Diverse and nonlinear influences of built environment factors on COVID-19 spread across townships in China at its initial stage. Scientific Reports 11(1): 12415.

19.

Zhang

(2023) Spatial and deep learning analyses of urban recovery from the impacts of COVID-19. Scientific Reports 13(1): 2447.

20.

Maithani

(2009) A neural network based urban growth model of an Indian city. Journal of the Indian Society of Remote Sensing 37(3): 363–376.

21.

Müürisepp

Järv

Sjöblom

, et al (2023) Segregation and the pandemic: The dynamics of daytime social diversity during COVID-19 in Greater Stockholm. Applied Geography 154: 102926.

22.

Nandam

Patel

(2021) A novel hybrid approach using SVM and spectral indices for enhanced land use land cover mapping of coastal urban plains. Geocarto International 37(16): 4714–4736.

23.

Niu

Wang

, et al (2020) Developing a deep learning framework with two-stage feature selection for multivariate financial time series forecasting. Expert Systems with Applications 148: 113237.

24.

Ogulenko

Benenson

Toger

, et al (2022) The fallacy of the closest antenna: Towards an adequate view of device location in the mobile network. Computers, Environment and Urban Systems 95: 101826.

25.

Park

Christensen

Lee

(2020) Unmanned aerial vehicles (UAVs) in behavior mapping: A case study of neighborhood parks. Urban Forestry & Urban Greening 52: 126693. DOI: 10.1016/j.ufug.2020.126693.

26.

Santiago-Iglesias

Romanillos

Sun

, et al (2024) Light in the darkness: Urban nightlife, analyzing the impact and recovery of COVID-19 using mobile phone data. Cities 153: 105276.

27.

Santos

Graw

Bonilla

(2019) A geographically weighted random forest approach for evaluate forest change drivers in the Northern Ecuadorian Amazon. Public Library of Science One 14(12): e0226224.

28.

Smith

LeSage

(2004) A bayesian probit model with spatial dependencies. Spatial and Spatiotemporal Econometrics 18: 127–160.

29.

Toger

Shuttleworth

Östh

(2020) How average is average? Temporal patterns in human behaviour as measured by mobile phone data – or why chose Thursdays. arXiv:2005.00137. DOI: 10.48550/arXiv.2005.00137.

30.

Vali

Comai

Matteucci

(2020) Deep learning for land use and land cover classification based on hyperspectral and multispectral earth observation data: A review. Remote Sensing 12(15): 2495.

31.

Wang

Huang

Yang

, et al (2022) Resilience and recovery of public transport use during COVID-19. npj Urban Sustainability 2(1): 18.

32.

Yuan

Liu

, et al (2024) The impact of COVID-19 on the jobs–housing dynamic balance: Empirical evidence from Wuhan between 2019, 2021, 2023. Land 13(8): Article 8. DOI: 10.3390/land13081299.

33.

Xiu

, et al (2021) Lockdown induced night-time light dynamics during the COVID-19 epidemic in global megacities. International Journal of Applied Earth Observation and Geoinformation 102: 102421.

34.

Yang

Wang

Zhang

, et al (2020) COVID-19, urbanization pattern and economic recovery: An analysis of Hubei, China. International Journal of Environmental Research and Public Health 17(24): 9577.

35.

You

Guan

(2024) Does self-containment of spatial scale and land use function contribute to mitigate urban heat island effects? Lessons from new towns in Shanghai. Land Use Policy 146: 107323.

36.

Zhang

H-D

Yang

J-M

, et al (2022) Spatiotemporal evolution of NO₂ diffusion in Beijing in response to COVID-19 lockdown using complex network. Chemosphere 293: 133631. DOI: 10.1016/j.chemosphere.2022.133631.