COVID-19 lockdown introduces human mobility pattern changes for both Guangdong-Hong Kong-Macao greater bay area and the San Francisco bay area

Abstract

In response to the coronavirus disease 2019 (COVID-19) pandemic, various countries have sought to control COVID-19 transmission by introducing non-pharmaceutical interventions. Restricting population mobility, by introducing social distancing, is one of the most widely used non-pharmaceutical interventions. Although similar population mobility restriction interventions were introduced, their impacts on COVID-19 transmission are often inconsistent across different regions and different time periods. These differences may provide critical information for tailoring COVID-19 control strategies. In this paper, anonymized high spatiotemporal resolution mobile-phone location data were employed to empirically analyze and quantify the impact of lockdowns on population mobility. Both the Guangdong-Hong Kong-Macao Greater Bay Area (GBA) in China and the San Francisco Bay Area (SBA) in the United States were studied. In response to the lockdowns, a general reduction in population mobility was observed, but the structural changes in mobility are very different between the two bays: 1) GBA mobility decreased by approximately 74.0–80.1% while the decrease of SBA was about 25.0–42.1%; 2) compared to SBA, the GBA had smoother volatility in daily volume during the lockdown. The volatility change indexes for GBA and SBA were 2.55% and 7.52%, respectively; 3) the effect of lockdown on short- to long-distance mobility was similar in GBA while the medium- and long-distance impact was more pronounced in SBA.

1. Introduction

Countries throughout the world enacted a series of containment strategies during the initial phase of the coronavirus disease 2019 (COVID-19) pandemic, aiming to control the spread of this disease (Fisher and Wilder-Smith, 2020, Sohrabi et al., 2020, Chung et al., 2021, Iftimie et al., 2021). These policies included traffic controls, self-quarantining, travel restrictions, and limitations of public activities such as closures of schools and businesses as well as suspended operations of industries or largely reduced production (Grépin et al., 2021, Gwee et al., 2021). People were ordered to stay at home, except for purchasing essential items or when seeking medical treatment (Gao et al., 2020). These intensive lockdown measures have led to a significant reduction in human mobility (Lai et al., 2020, Zhou et al., 2020), which had been identified as a major driver of COVID-19 transmission (Jia et al., 2020, Yang et al., 2020).

However, it is still a controversially discussed topic whether changes of human mobility are strongly associated with COVID-19 transmission. For example, while two regions implemented similar human mobility restriction interventions, it was observed that the effectiveness of epidemic control between both differed significantly. In addition, a study conducted in the US found a strong correlation between human mobility volumes and COVID-19 cases in the first wave of the pandemic (January to April 2020). However, the correlation was significantly weakened if the analysis was extended to September 2020 (Badr et al., 2020, Gatalo et al., 2021, Badr and Gardner, 2021). The driver for these inconsistent analytical results still remains unknown.

The authors suggest that one key driver for this inconsistency is the granularity of employed mobility analyses. Previous studies generally focused on changes of mobility volume, while fine-grained mobility indicators (e.g., distance and stability) were not available. This may introduce differences in structural changes of human mobility between different regions or different time periods. This limitation provides a major motivation for the present research, where structural changes are quantified in daily mobility (overall, intra- and inter-city mobility volume, and distance-dependent mobility volume), travel distance, and stability for the first wave of the COVID-19 epidemic in the Guangdong-Hong Kong-Macao Greater Bay Area (GBA) and the San Francisco Bay Area (SBA). These four indicators are built to provide an assessment of the extent to which people in both the GBA and SBA have reduced their mobility and physical proximity during the COVID-19 pandemic. Anonymized, aggregated mobile phone positioning data were employed to build intercity mobility networks and explore spatiotemporal mobility patterns using truncated singular value decomposition (TSVD). Analyzing these patterns facilitates a deeper understanding of changes in human behavior, which provides effective information for fighting future pandemic waves, tailoring prevention and control strategies, and managing human mobility towards controlling the continued spread of COVID-19 (Chinazzi et al., 2020, Sills et al., 2020, Xu et al., 2021).

2. Related work

2.1. Understanding human mobility patterns

Human mobility is an essential component of human development (De Haas and Rodríguez, 2010). Human mobility analysis strives to comprehend the intrinsic properties of human movements and the mechanisms behind observed patterns (Szell et al., 2012, Xu et al., 2018). Models of human mobility indicate a high degree of regularity instead of randomness in human population movements (González et al., 2008, Gao et al., 2020) and thus enable a certain level of predictability (Badr et al., 2020, Xiong et al., 2020) at multiple spatiotemporal scales (Simini et al., 2012, Pappalardo et al., 2015, Yan et al., 2017). These patterns of movements change with holidays (Deville et al., 2014) as well as emergencies (Wesolowski et al., 2015). When these two factors occur simultaneously, the associated large increases in the volume of human mobility greatly accelerate the spread of infectious diseases (Lai et al., 2020). Human mobility and emergencies impose a coupled effect on disease transmission. Therefore, controlling the volume of human mobility in the early stages of an outbreak can more effectively delay the spread of diseases (Kraemer et al., 2020, Zhou et al., 2020).

2.2. Human mobility impacts of COVID-19

Recent advancements in mobile phone location detection (Wang et al., 2020) and location-aware technologies (Wang et al., 2019) have provided numerous new datasets. These hold the potential to evaluate human movement dynamics at unprecedented geographical and temporal resolutions and scales (Blondel et al., 2015), often, in nearly real-time (Pullano et al., 2020). Mobility data have been used extensively during the COVID-19 pandemic to better comprehend epidemic transmission between populations (Oliver et al., 2020, Xiong et al., 2020), identify transmission hotspots (Chang et al., 2021), and guide policy interventions (Lucchini et al., 2021). For instance, Guan et al. (2021) combined mobility data and case data comprising travel histories to assess the influence of human mobility on the COVID-19 pandemic in Brazil. Beria and Lunkar (2021) used Facebook data and found that lockdowns in Italy resulted in a near-zero reduction in both local and national mobility. Jia et al. (2020) conducted a comprehensive analysis in China using a mix of mobile phone data and COVID-19 case data. Their findings showed that the frequency and geographic distributions of illnesses are well predictable by the pattern of population outflow. Klein et al. (2020) used Cubeiq mobility data to explore mobility trends in the USA. To evaluate the degree of social separation and address concerns about the course of the pandemic, they quantified mobility loss at the macro, meso, and microscopic levels. Saha and Chouhan (2021) used Google community mobility data for India and found a dramatic decrease in mobility patterns, with the exception of residential mobility.

In general, previous studies focused on analyzing human mobility volume and its impacts on COVID-19. However, fine-grained mobility indicators (e.g., distance and stability) are still missing, which may result in different structural changes of human mobility between different regions or time periods. Therefore, this research proposes four fine-grained human mobility index changes that were introduced by the COVID-19 imposed lockdowns. A systematic understanding of these changes is critical for assessing the effectiveness of non-pharmaceutical interventions (Prem et al., 2020) and minimizing the spread of the pandemic via advanced preparation measures (Iacus et al., 2020, Kraemer et al., 2020).

3. Research area and research data

3.1. Study area

This study employed data on two bay areas, the GBA and the SBA. Both bay regions have a comparable number of cities, as well as a large concentration of high-tech enterprises, financial services, cultural industries, and educational resources. Moreover, both bay areas have high population density and mobility, identifying them as the most representative urban clusters in China and the USA, respectively. As shown in Fig. 1 (A), the SBA is located on the west coast of the USA and includes nine counties: San Francisco, Alameda, San Mateo, Sonoma, Solano, Napa, Marin, Santa Clara, and Contra Costa. The GBA is located in southern China, adjacent to the South Pacific Ocean, and includes nine prefecture-level cities, i.e., Shenzhen, Zhaoqing, Jiangmen, Huizhou, Dongguan, Foshan, Zhuhai, Zhongshan, and Guangzhou, as shown in Fig. 1(B). While the GBA also comprises the two special administrative regions of Macau and Hong Kong, because of data unavailability, these regions were excluded from this study. Table 1 provides more information on the both regions.

Fig. 1.

Geographical location and administrative divisions of the two bay areas: (A) Guangdong-Hong Kong-Macao Greater Bay Area (GBA), and (B) San Francisco Bay Area (SBA).

Table 1.

Basic Information of SBA and GBA as of the end of 2020.

	SBA	GBA
Latitude (N)	36°56′–38°47′	21°32′–24°26′
Longitude (E)	121°39′–123°33′	111°20′–115°24′
Contained cities	9 counties	11 cities
Population	7.74 million	86.17 million
Area (km2)	1.8 × 104	5.6 × 104
Population density	427 per/km2	1182 per/km2
Percentage of population size	2.30%	4.80%
Percentage of land area	0.19%	0.59%
GDP(US dollars)	0.99 trillion	1.67 trillion
Percentage of GDP	9.3%	12.5%

Note: Data from HKTDC (2020).

3.2. Dataset and data preprocessing

The mobile phone data for the GBA is provided by one of the three leading mobile phone service providers in Guangdong province, covering approximately 20% of the total GBA population. China Unicom cell phone operators developed relevant GBA-wide population coverage models and machine learning algorithms, and used Cellular Signaling data combined with census data to derive the total population movement in the GBA. They used various parameters extracted from the call ratio, user age, gender, and other demographic structures with other operators. The authors cross-validated the cell phone imputed population movement data with the open dataset (Baidu index data), which had a correlation of 0.82. This study employs anonymized aggregated location data from January 1 to March 15, 2020, the interval that represents the entire cycle of the first wave of emergence, prevention, and control of COVID-19 in China.

The SBA data was acquired from the GeoDS Lab at the University of Wisconsin-Madison, which employed cell phone location data provided by SafeGraph for approximately 10% of the total U.S. population. These data were combined with American Community Survey demographic data with cell phone visitor patterns to infer a dynamic overall Origin-Destination(OD) population mobility matrix. This population mobility dataset has a high correlation (R² > 0.92) with publicly available data sources and is highly reliable after having been validated with three complementary methods. This paper uses anonymized aggregated location statistics from March 1 to May 14, 2020, during the first wave of COVID-19 in the USA. The period was kept consistent with the GBA, totaling 75 days.

Based on city-level anonymous aggregated human mobility big data, the total volume of daily population movement was aggregated to construct a daily flow matrix $\mathcal{F}(\mathbf{t})$. F_ij(t) quantifies the total population movement from city i to city j on a given date t, where F_ii(t) is the intra-city population flow of city i. Because the movement of people is omnidirectional, the total movement from city i to city j is not equal to the total movement from city j to city i. $\mathcal{F}(\mathbf{t})$ characterizes the amount of population movement between and within cities, and it contains a typical commuting pattern and micro-scale movement of people (Badr et al., 2020). The baseline dates were the normal days in 2020, defined as the days from January 1 to January 7, 2020 for GBA, and the dates from March 1 to March 7, 2020 for SBA, when travel patterns were stable.

4. Methodology

Four evaluation metrics were constructed to quantify structural changes in human mobility during the lockdown periods. Then, the TSVD method was utilized to investigate spatiotemporal mobility patterns.

4.1. Mobility change Index

To quantify overall mobility trends in the bay areas, this paper focuses on the total amount of trips $N(t)=\sum _{i,j=1}^m{F}_{\mathit{ij}}\left(t\right)$ on date t, where m is the number of cities in the bay area. To explore changes in intracity and intercity human mobility during the pandemic, $N_{\mathit{intra}}^{\left(t\right)}$ and $N_{\mathit{inter}}^{(t)}$, the identical procedure was used and the total volume of intracity and intercity human mobility in each bay area was summed up at day t, i.e.,$N_{\mathit{intra}}^{\left(t\right)}={\sum }_{i=1}^m{F}_{\mathit{ii}}\left(t\right)$, $N_{\mathit{inter}}^{\left(t\right)}={\sum }_{i\ne j}^m{F}_{\mathit{ij}}\left(t\right)$, where m is the number of cities in the bay area. To explore human mobility changes during the lockdown, the Mobility Change Index (MCI) was defined for each day t as MCI(t). For each given day t during the pandemic, the mobility change MCI(t) was computed by comparing the total volume of travel N(t) to the expected volume of travel N₀(τ₀) during the baseline as.

$$\mathrm{M}\mathrm{C}\mathrm{I}\left(\mathrm{t}\right)=\frac{N\left(t\right)}{{\mathrm{N}}_0\left({\tau }_0\right)}-1$$

Using this function, MCI(t) values>0 and less than 0 indicate an increase and decrease in the volume travel compared to baseline, respectively. Larger absolute values indicate stronger increases or decreases. MCI(t) captures the change in the volume of individual movements per day in each city, compared to normal behavioral patterns (i.e., before COVID-19).

4.2. Distance change Index

To measure the changes in individual travel distance in Bay Area during the pandemic, we defined the Distance Change Index (DCI) for each day t as DCI(t). Because of regulatory and ethical requirements for the use of mobile phone data, aggregated-scale anonymized data were used, where individual-scale travel behavior could not be extracted. The geodesic distance (Karney, 2013) d_ij between the center of mass of spatial cell i and j was used as an estimate of travel distance. The weighted average distance of travel at day t is.

$$\mathrm{D}(\mathrm{t})=\frac{{\sum }_{i,j=1}^m{N}_i(t)·d_{\mathit{ij}}(t)}{N(t)}$$

where N_i(t) is the total volume of travel in spatial cell i on day t, and N(t) is the same as the previous section but only the travel that originated in the spatial cell i was calculated. D ₀(τ ₀) is the weighted average distance of trips on the baseline. DCI is calculated as follows:

$$\mathrm{D}\mathrm{C}\mathrm{I}(\mathrm{t})=\frac{D(t)}{D_0({\tau }_0)}$$

using this function, DCI(t) values of 0, 0.5, and 1.0 signified no trips, half the travel distance relative to baseline, and no change compared with the baseline period, respectively. DCI(t) shows the changes in the individual travels distance in each region every day, compared to normal behavioral patterns (i.e., before COVID-19).

4.3. Distance-dependent mobility changes

Distance-dependent Mobility Changes (DDMC) calculate the changes of mobility volume based on different mobility distances (e.g., ≤10 km, >10 km ≤ 50 km, >50 km ≤ 100 km, >100 km). When calculating distance-dependent mobility changes, the previous section was duplicated. The volume of travel in the distance range d_ij is$\text{MD}\left(\text{t}\right)={\sum }_{\left(i,j\right)\in d_{\mathit{ij}}}{F}_{\mathit{ij}}\left(t\right)$, and MMCD is calculated as.

$$\mathrm{M}\mathrm{M}\mathrm{C}\mathrm{D}(\mathrm{t})=\frac{\mathit{MD}(t)}{\mathit{MD}({\tau }_0)}-1$$

where MD(τ ₀) denotes the set of all spatial cell pairings (i, j) whose distances fall within d_ij on the baseline. Using this function, $\mathrm{M}\mathrm{M}\mathrm{C}\mathrm{D}(\mathrm{t})$ values >0 and less than 0 indicate an increase and decrease in the volume travel for different travel distances compared to the baseline, respectively. Larger absolute values indicate more increases or decreases. $\mathrm{M}\mathrm{M}\mathrm{C}\mathrm{D}(\mathrm{t})$ captures the change in the volume of travel at different trip distances per day in each city, related to normal behavioral patterns (i.e., before COVID-19).

4.4. Volatility Change Index

The Volatility Change Index (VCI) is defined to measure the stability of human mobility during these phases of the pandemic. Because the MCI can only reflect population movement changes at a macro level, it cannot reflect the stability of population flows at a given time. Therefore, the VCI is established as an indicator for measuring the stability of population movement.

$$\mathrm{\Delta }{\mathrm{N}}_i(t)=ln(\frac{N_i(t)}{N_i(t-1)})$$

$$\mathrm{\Delta }{\overline{\mathrm{N}}}_i(\tau )=\frac{{\sum }_{\tau }\mathrm{\Delta }{\mathrm{N}}_i(t)}{\tau (\tau =7)}$$

$${\mathit{VCI}}^{\left(i\right)}\left(\tau \right)=\sqrt{\frac{\sum {\left(\mathrm{\Delta }{\mathrm{N}}_i\left(t\right)-\mathrm{\Delta }{\mathrm{N}}_i\left(\tau \right)\right)}^2}{\tau -1}}\times 100\%$$

where ${\mathrm{N}}_i\left(t\right)$ represents the total travel volume of city i on day t. $\mathrm{\Delta }{\mathrm{N}}_i\left(t\right)$ denotes the natural logarithmic difference between the total number of population movements on day t and the previous day. $\mathrm{\Delta }{\overline{\mathrm{N}}}_i(\tau )$ denotes the average population movement of city i during the period $\mathrm{\tau }$, where $\mathrm{\tau }$ encompasses the last 7 days. ${\mathit{VCI}}^{\left(i\right)}\left(\tau \right)$ is the VCI of city i in period $\mathrm{\tau }$. Using this function, ${\mathit{VCI}}^{\left(i\right)}\left(\tau \right)$ values indicate the stability of population movements, with larger values indicating less stable population movements.

4.5. Spatiotemporal matrix decomposition using TSVD

Powerful matrix analysis techniques include principal component analysis (PCA) (Sun and Axhausen, 2016), singular value decomposition (SVD) (Chen et al., 2018), and non-negative matrix decomposition (Huang et al., 2013). In most cases, PCA problems can be transformed into SVD problems, and using SVD is usually more stable than using PCA directly as SVD avoids part of the accuracy loss in covariance calculations (Lipovetsky, 2009; Li et al., 2021). In addition, SVD decomposes a data matrix into uncorrelated variables, revealing intrinsic structures in multi-dimensional data (Jolliffe and Cadima, 2016). Different from other matrix decomposition methods, SVD can be applied to any matrix (Bac and Zinovyev, 2020). By decomposing a matrix with SVD, PCA results can be obtained in both directions (row direction and column direction), corresponding to important temporal and spatial features in specific problems (Attachment 1). The TSVD provides a formal method for the rank-restricted optimum approximation of matrices by substituting the lowest singular values with zeros in the SVD of a matrix (Zekri et al., 2019, Vu et al., 2021). In this paper, the TSVD method is used to identify the main potential characteristics of each dimension of the spatiotemporal matrix (Fig. 2 ).

Fig. 2.

Illustration of the decomposition of the spatiotemporal matrix M by truncated singular value decomposition (TSVD).

The spatiotemporal matrix M represents the set of intercity (county) population mobility types in a bay area during the pandemic. $m_{\mathit{ij}(i\ne j)}$ represents the total number of population movements between the origin city i and the arrival city j on date t. Because human mobility is directional, $m_{\mathit{ij}}$ is not equal to $m_{\mathit{ji}}$, which transforms the bay area intercity OD into a $1\times \mathrm{n}$ row vector on the specified date t, stacking the s days vertically into an $\mathrm{s}\times \mathrm{n}$-dimensional matrix M. If the rank of M is r and r < rank (M), then.

$$M_r={\mathit{USV}}^T={\sum }_{h=1}^r{\delta }_h{u}_h{v}_h^T$$

is a r-truncated singular value decomposition of M. U is an $\mathrm{s}\times \mathrm{r}$ orthogonal matrix while $u_h$ denotes the h-th column of U. S is a $r\times r$ diagonal matrix with descending singular values, the h-th singular value is ${\delta }_h$, i.e., ${\delta }_1\ge {\delta }_2···\ge {\delta }_r$. $V^T$ an $\mathrm{r}\times \mathrm{n}$ orthogonal matrix and $v_h^T$ denotes the h-th row of $V^T$.

After TSVD decomposition of matrix $M_r$, the left singular matrix U represents the changing patterns in the temporal dimension, while the right singular matrix $V^T$ represents the changing patterns in the spatial dimension. $u_h$ denotes the temporal distribution of the h-th mobility pattern, $v_h^T$ represents the spatial distribution of the h-th mobility pattern and ${\delta }_h$ represents the importance of the h-th mobility pattern. If the top k (1 ≤ k ≤ r) greatest singular values are retained and the remaining smaller singular values are removed, the most essential k terms will be retained and describe the most important types of human mobility during the pandemic.

5. Results

5.1. Mobility structural changes in GBA and SBA

See Fig. 3 .

Fig. 3.

Spatial and temporal changes in mobility during the COVID-19 pandemic. (A) Changes in total, intercity and intracity mobility in GBA, relative to baseline. The points indicate the calculated values and the lines indicate the generalized additive model (GAM) fitting curves. The two vertical dashed lines indicate the policy implementation time points (January 23, 2020 and February 24, 2020). (B) and (E) intercity and intracity mobility changes in GBA and SBA during Pre-Lockdown, Lockdown and Post-Lockdown for three time periods respectively, based on policy implementation dates. (C) and (F) Correlation of the volume of changes in intercity and intracity mobility levels in GBA and SBA, respectively. (D) Changes in total, intercity and intracity mobility in SBA, relative to baseline. The points indicate the calculated values and the lines indicate the GAM fitting curves. The two vertical dashed lines indicate policy implementation time points (March 19, 2020 and May 4, 2020). (G) and (H) Spatial changes of Mobility Change Index (MCI) in GBA and SBA during the week before the Pre-Lockdown period, the second week of the Lockdown period, and the first week of the Post-Lockdown period, respectively.

5.1.1. Mobility temporal changes

As shown in Fig. 3(A), GBA mobility declines sharply in late-January, which coincides with the implementation of restrictive measures. The MCI of GBA decreased by about 74.0–80.1% compared with baseline, then retains this low level for about three weeks, implying that because of restrictive policies, mobility followed a steady upward trend. The volume of change in mobility for intracity and intercity is consistent with the magnitude of change in MCI. Similarly, the SBA mobility declined in mid-March and eventually became smaller than that of GBA. The MCI of SBA remained generally stable over the lockdown period, with a more apparent mobility characteristic of a weekly change cycle. The MCI decreased by about 25.1–42.1% compared with the daily pattern. Unlike the GBA, the total volume of intercity travel declined more than MCI and intracity in the SBA (Fig. 3(D)). A 7-days time growth node was used and the MCI of GBA and SBA decreased by 62.7–78.2% and 29.9–46.3%, respectively, during the lockdown period (Table 2, Table 3 ).

Table 2.

Weekly changes in MCI during GBA activated first-level public health emergency response.

	Week 4	Week 5	Week 6	Week 7	Week 8	Average
Dongguan	−81.29%	−83.91%	−80.53%	−75.47%	−66.43%	−77.53%
Zhongshan	−74.83%	−82.06%	−83.63%	−81.45%	−67.36%	−77.87%
Foshan	−74.28%	−81.71%	−79.21%	−74.44%	−62.36%	−74.40%
Guangzhou	−75.67%	−82.07%	−79.35%	−75.05%	−68.58%	−76.14%
Huizhou	−68.42%	−75.46%	−74.96%	−73.78%	−64.62%	−71.45%
Jiangmen	−47.62%	−69.51%	−71.19%	−69.71%	−55.51%	−62.71%
Shenzhen	−79.50%	−83.47%	−82.32%	−78.05%	−67.64%	−78.20%
Zhuhai	−69.37%	−80.72%	−80.34%	−74.62%	−66.64%	−74.34%
Zhaoqing	−43.83%	−65.43%	−73.84%	−73.85%	−57.35%	−62.86%

Table 3.

Weekly changes in MCI during the SBA “Stay at Home” (i.e., Lockdown) order.

	Week 12	Week 13	Week 14	Week 15	Week 16	Average
Alameda	−39.30%	−38.91%	−40.90%	−39.68%	−38.07%	−39.37%
Contra Costa	−35.41%	−34.69%	−37.20%	−34.61%	–32.61%	−34.91%
Marin	−38.85%	−38.53%	−40.35%	−37.72%	−36.80%	−38.45%
Napa	−35.57%	−36.34%	−38.25%	−35.40%	–33.35%	−35.78%
San Francisco	−45.35%	−45.92%	−47.92%	−46.44%	−45.76%	−46.28%
San Mateo	−42.45%	−42.79%	−45.57%	−43.46%	−41.82%	−43.22%
Santa Clara	−40.54%	−40.52%	−42.31%	−40.49%	−38.82%	−40.53%
Solano	−30.20%	−29.96%	–32.68%	−29.29%	−27.11%	−29.85%
Sonoma	–33.13%	–33.86%	−35.60%	–32.46%	−30.35%	–33.08%

Regarding the intracity and intercity human mobility changes, correlations of 0.98 and 0.94 were found between GBA and SBA (Fig. 3(C) and Fig. 3(F)), respectively, with consistent changes in both bays. This was divided into three intervals based on the duration of the lockdown period: Pre-Lockdown, Lockdown, and Post-Lockdown. In the GBA, as shown in Fig. 3(B), in the Pre-Lockdown period, the intracity and intercity mobility declined by 17.2% and 6.2%, respectively. During the Lockdown period, both intracity and intercity human mobility dropped significantly, by 74.0% and 77.1%, respectively. In the Post-lockdown period, the total mobility gradually increased, decreasing by about 42.5% and 41.1% for intracity and intercity, respectively, compared with baseline, but still not reaching Pre-lockdown levels. The overall changes in SBA and GBA are different. As shown in Fig. 3(E), in the SBA, the change in intercity mobility is about 2.08 times greater than that in intracity mobility. During the Lockdown periods, intracity and intercity mobility changes decreased by 0.28 and 0.59, respectively, both of which are smaller than in the GBA, especially the intracity mobility change, which was only 37.8% of that in the GBA.

5.1.2. Mobility spatial changes

In GBA, the change in mobility is spatially heterogeneous. As shown in Fig. 3 (G), except for Zhaoqing and Jiangmen, all other seven cities decreased below 20% one week before the Lockdown. The possible reason for the this “abnormal fluctuation” in these two cities is that the pandemic outbreak coincided with the Chinese traditional Spring Festival, and Zhaoqing and Jiangmen are the major labor exporters of GBA. Therefore, this fluctuation may be related to the large number of people returning home during the Chunyun period. During the second week of the Lockdown period, all cities of the GBA experienced a dramatic decline in human mobility, all exceeding 65% or even more, with the mobility volumes of Dongguan, Foshan, Guangzhou, Shenzhen, Zhongshan, and Zhuhai declining strongest, by over 80%. In the Post-Lockdown period, the overall MCI of GBA remained consistent at around 50%, with no significant difference among cities. In SBA, the spatial change in mobility is more similar. As shown in Fig. 3(H), one week before the Lockdown, the change in mobility only decreased by about 10–20%, with the city of San Francisco decreasing the most, by about 25%. The second week of the Lockdown was the week when mobility decreased the most (by about 30–39%) because of the implementation of the “Stay at Home” order. In the Post-Lockdown period, although the government reopened the economy, there was no significant rebound in mobility change, with approximately 24–43% less than the normal baseline. Throughout the study interval, mobility declined slightly more in the southern part of the SBA than in the northern part, and the MCI in San Francisco City, Peninsula, South Bay, and East Bay regions are lower than in the North Bay. The SBA formed a south-low and north-high spatial distribution structure.

5.1.3. Mobility cycle changes magnitude

Fig. 4(A) and (B) show the VCI of GBA and SBA during the Lockdown period, respectively, and the areas of the circles indicates the VCI values. The results show that, during the Lockdown period, the fluctuation of GBA was much smaller than that of the SBA. The first two weeks of VCI show a cyclical change pattern, and the weekday VCI is larger than that of the weekend. After the third week, the VCI decreases to the lowest level and remains at this level for two or three weeks, i.e., the total volume of daily mobility is more constant and the changes tend to be stable during this phase. Compared with GBA, SBA maintains a high VCI level throughout the Lockdown period, i.e., the total daily mobility volume is more variable. The average VCI of GBA and SBA during the Lockdown period is 2.55% and 7.52%, respectively, which is a difference of about 3 times. From geospatial distribution perspective, there is no significant difference between VCI between cities in the two regions and the change is consistent.

Fig. 4.

Volatility variation of GBA and SBA during Lockdown and distance-dependent mobility changes. (A) and (B) The Volatility Change Indexes (VCI) of GBA and SBA respectively. The color represents different cities, and the size of the circle areas represents the VCI value. (C) and (D) Variation of GBA and SBA mobility at different distance ranges, respectively (points are original values and curves are generalized additive model fitting curves). (E) and (F) Comparison of distance change, average mobility change, and short-long distance mobility difference for GBA and SBA, respectively. Green denotes the overall average distance traveled in the bay areas (DCI), while blue is the average number of trips per day compared to baseline. Orange indicates the difference in mobility changes between short-distance ($d_{\mathit{ij}}\le 10$ km) and long-distance ($d_{\mathit{ij}}>10$ km) travel, which is a useful indication for atypical mobility patterns.

5.1.4. Distance dependence on mobility reduction

DDMC are explored in both bay areas. Except for fluctuations in mobility volumes for travel distances exceeding 100 km, the decreasing trend is the same for short- or medium-distance trips in the GBA. Moreover, the mobility volume size remains low during the Lockdown (Fig. 4C). The overall DCI of GBA also decreases to the lowest level during the Lockdown period, decreasing 17.44% compared to the baseline. The DCI is almost 0 during the Post-Lockdown period, i.e., the GBA average travel distance gradually returns to a normal level. Furthermore, a difference in distance-dependent mobility change was found between short-distance and long-distance mobility. In the second half of the Lockdown period, the reduction in long-distance mobility is more significant than that of short-distance mobility, and gradually recovered. The reduction of the short-distance mobility volume is stronger in other periods. It is possible that the end of the Chinese New Year holiday contributed to the recovery of inter-city travel and the increase of long distances travels (from family reunion back to work). In the Post-Lockdown period, the DDMC is around the 0 value, i.e., the effect of distance on travel is weaker. However, the mobility volume does not return to normal levels, and about a 33% absolute gap remains (Fig. 4E). During the Lockdown period, different effects of SBA and GBA on DDMC were found, and SBA decreased stronger for long-distance mobility than for short-distance mobility (Fig. 4D). This coincides with policies: long-distance travel bans and cancellation of major events led to weaker demand for long-distance travel. The DCI of the SBA decreased by 9.15% on average, following a general decreasing trend, including a minimum value of 36.6% on May 1. The DCI drops below 0 after 10 days of the Lockdown period, and the weekend travel distance is much greater than that of weekdays. During the Post-Lockdown period, the DCI showed a further reduction, but the average mobility volumes changed inconspicuously, still showing a 49% stable difference (Fig. 4F).

5.2. Spatiotemporal mobility pattern discovery

The TSVD algorithm was used to decompose the human mobility matrix of the two bay areas during the pandemic to reduce the dimensionality and identify mobility patterns. The singular value ${\mathrm{\delta }}_{\mathrm{h}}$ is an index to measure the importance of the intercity (i.e., county) human mobility patterns. The larger the singular value, the more original information is represented by the corresponding mobility pattern. The normalized singular value distributions of GBA and SBA are displayed in Fig. 5 (A) and (B), respectively. The singular value of M decreases sharply and the first value is much larger than all other values. The first three largest singular values of GBA and SBA represent 86.3% and 82.3% of M original information, respectively.

Fig. 5.

Distribution of Normalized Singular Values of M for(A) GBA and (B) SBA.

Based on this, the three most important intercity (county) human movement patterns from each bay area spatiotemporal matrix M are identified. In the result, Fig. 6, Fig. 7 (A), (C), and (E) show the temporal change in this mobility pattern, where the absolute value represents the relative change of human mobility, and positive and negative values indicate the fluctuation direction of the human movement on the y-axis. Fig. 6, Fig. 7 (D), (E), and (F) show the spatial changes in this mobility pattern where red and green colors represent positive and negative fluctuations in human mobility, respectively; the darker the color, the more significant the fluctuation. Combining temporal and spatial flows can, explain the specific meaning of each pattern: if within a certain period, the temporal flow of significant positive fluctuations, and the spatial flow are positive, i.e., the temporal and spatial flows fluctuate in the same direction, total human mobility increases significantly. On the contrary, if temporal and spatial flows show inverse fluctuations, the total volume of population flows decreases significantly.

Fig. 6.

Human mobility pattern of GBA. (A) Temporal flows of pattern I. The green vertical line indicates the first confirmed case of COVID-19 infection in Guangdong province on January 19. The red vertical line indicates the activation of the first-level public health emergency response in Guangdong province on January 23. The blue vertical line indicates the lowering of the emergency response level, and the release of traffic control on February 24. (B) Intercity spatial interaction of the pattern I. (C) Temporal flows of pattern II. The green vertical line indicates the start of the Spring Festival on January 10. The red vertical line indicates the first day of the Spring Festival holiday on January 24. The blue vertical line indicates that the government downgraded the response level of public health emergencies on February 24. (D) Intercity spatial interaction of the pattern II. (E) Temporal flows of the pattern III. The green vertical line indicates the first day of the Spring Festival holiday on January 24. The red vertical line indicates that the State Council announces the extension of the Chinese New Year holiday on January 27. The three blue vertical lines indicate the three wave peaks during the Spring Festival holiday. (F) Intercity spatial interaction of pattern III.

Fig. 7.

Human mobility pattern of SBA. (A) Temporal flows of pattern I. The green vertical line indicates the first regional order issued in Puerto Rico on March 15. The red vertical line indicates that the “Stay at Home” order was enacted and implemented in California on March 19. The blue vertical line indicates that the government relaxed the lockdown and reopened the economy on May 4. (B) Intercity spatial interaction of pattern I. (C) Temporal flows of pattern II. The green vertical line indicates the last working day of the week before the government implemented the lockdown on March 13. The red vertical line indicates that the “Stay at Home” order was enacted and implemented in California on March 19. The blue vertical line indicates the reopening of businesses and schools on May 4 for resumption of work, schooling, and production. (D) Intercity spatial interaction of pattern II. (E) Temporal flows of pattern III. The two green vertical lines indicate the weekend before the government implemented the lockdown on March 13 and 14. The red vertical line indicates that the “Stay at Home” order was enacted and implemented in California on March 19. The two blue vertical lines indicate the first weekend of the reopening of the economy on May 9 and 10. (F) Intercity spatial interaction of pattern III.

5.2.1. Identification of human mobility patterns in GBA

Pattern I represents the daily population movement characteristics. As shown in Fig. 6A, using January 19, 2020, as the time boundary, the fluctuation of temporal distribution flow is relatively stable both before and after phases. Guangdong Province activated its First-Level Public Health Emergency Response on January 23, 2020. In GBA, the COVID-19 pandemic coincided with the Chinese New Year, a traditional Chinese holiday. Under the combined influence of the Chinese New Year long holiday effect and the public health policy effect, the intercity characteristic value decreased by 90.5% within a short period. From the perspective of spatial flow (Fig. 6B), the cities of Guangzhou-Foshan and Shenzhen-Dongguan have the closest population flow interactions, and the daily population flow pattern of the overall GBA reflects a more apparent geographical proximity effect.

Pattern II represents the intercity homecoming characteristics. As shown in Fig. 6C since the beginning of the Spring Festival on January 10, 2020, Pattern II retained positive values, which declined gradually after reaching a positive maximum on January 18, and returning to negative values two days after the start of the Spring Festival holiday. Significant positive fluctuation exists in the temporal distribution flow between in January 10–26, with a more pronounced increase in population movement because of homecoming. On February 24, the GBA public health emergency response downgraded to two levels, after which the negative fluctuation of the inter-city population movement scale increased gradually. From the perspective of spatial flow (Fig. 6D), the pre-holiday homecoming characteristics mainly showed population inflow from developed cities such as Guangzhou, Shenzhen and Foshan to less developed cities such as Zhaoqing and Huizhou, and the population migration trajectory followed a strip-like distribution. Guangzhou and Shenzhen, both international comprehensive transportation hub cities, have the most significant scale of population movement in the pre-holiday homecoming feature. After the Spring Festival, the scale of returning population movement was not as large as before the festival, With the response level adjusted to two levels, GBA local homecoming features are more apparent.

Pattern III represents the return-to-work (RTW) characteristics. As shown in Fig. 6E, because of the strict traffic control measures at the beginning of the COVID-19 pandemic and the uncertain time for enterprises RTW and production, after the start of the Chinese New Year holiday on 24 January, the pattern III temporal flow shows three more obvious positive fluctuations. The positive maxima of the three fluctuations were reached on 29 January, 8 February, and 23 February. The State Council issued an order on January 27 to extend the Spring Festival holiday in 2020, and the first post-holiday RTW peak occurred two days later. The second RTW peak occurred on the last weekend before enterprises resumed work and production in Guangdong Province. The third RTW peak occurred the day before Guangdong Province downgraded its public health emergency response level. All three peaks occurred earlier or later than the promulgation of the policy, which may be closely related to the staggered RTW of certain enterprises and whether the local pandemic risk was deemed controllable. Combination of the spatiotemporal flows to analyze the human mobility characteristics of RTW showed that the direction of population movement is basically opposite to pattern II (Fig. 6F).

5.2.2. Identification of human mobility patterns in SBA

Pattern I represents the daily population movement characteristics of the SBA. As shown in Fig. 7A, California enacted and implemented a “Stay at Home” order on March 19, 2020. Bounded by March 13, the temporal flows decreased from 0.22 in the previous period to 0.05 in the following period. The population mobility scale decreased earlier than the lockdown order implementation date, suggesting that certain public health policies may have influenced behavior in other areas. The SBA lockdown order has a lag effect of about one week, and the population flow characteristic value between counties decreased by 77.3% over a short time. From the perspective of spatial distribution flow (Fig. 7B), the population flows between counties show apparent geographical proximity effects, with more significant intercounty interactions in the southern bay area and weaker links between the north bay.

Pattern II represents the intercounty weekday travel characteristics of the SBA. As shown in Fig. 7C, on March 19, 2020, the California Executive Order and Public Health Order directed all Californians to stay at home unless they are traveling to perform essential work or shop for essential needs. Pattern II temporal flows show that the eigenvalues retain positive fluctuations for the two weekdays before the implementation of the lockdown order. The eigenvalues drop to negative values, but around the value of 0, the negative fluctuation is less significant. From the perspective of spatial distribution flow, as shown in Fig. 7D, the spatial flows are mainly to San Francisco city, while intercounty interactions in the north bay region are not significant. The Peninsula and the San Francisco area have the highest residential densities. Before the lockdown, population movement showed a convergent distribution centered on San Francisco city, with a sharp decrease in the size of human flows towards San Francisco city as destination, but an increase in the size of the origins. This may be related to government-mandated closure of schools and businesses, which decreased the demand for people to travel on workdays, and increased demand for remote work and learning from home.

Pattern III represents the intercounty weekend travel characteristics of the SBA. As shown in Fig. 7E, starting from the weekend of the last week when the lockdown order was implemented, positive weekly fluctuations in the eigenvalues gradually decreased and reached their lowest values in the third week of the Lockdown Period, after which the weekend travel eigenvalues remained largely stable. On May 4, the state modified the mandatory closure order and reopened the economy. Intercounty travel characteristic values increased significantly in the first weekend after this reopening and returned to the levels during Lockdown. Combined with the spatial flows, as shown in Fig. 7F, all cities interacted with each other in terms of population flows. Compared to Pattern II, weekend travel connections in the north and south of the bay area were significantly stronger, and people’s demand for long-distance travel increased.

6. Conclusion and limitations

In this study, structural changes in human mobility during the first wave of the COVID-19 pandemic were investigated for both GBA and SBA. Furthermore, the intrinsic structure and mobility patterns of population movement at the city scale were explored to gain a dynamic understanding of the spatial and temporal evolution patterns of human mobility across different periods. The major conclusions are summarized in the following: first, during the COVID-19 lockdown, mobility decreased dramatically in GBA and SBA, with overall GBA mobility decreased by approximately 74.0–80.1% and SBA mobility decreased by about 25.0–42.1% compared to baseline. Intra- and inter-city mobility decreased by 74% and 77.0% for GBA, and by 28.0% and 59.0% for SBA, respectively, with greater mobility declines observed for GBA than for SBA. Second, the average VCI of GBA and SBA during the Lockdown period were 2.55% and 7.52%, respectively, with a difference of about 3 times. The volatility in daily volume of GBA mobility was smoother during the lockdown. Third, the effect of lockdown on short- to long-distance mobility was similar in GBA, with all types of mobility staying low. Medium- and long-distance impacts are most pronounced in SBA. In addition, during the Post-Lockdown period, the DCI and DDMC of GBA recovered to about the value 0, but the overall average mobility still had about a 33% gap over baseline. The DCI decreased further in SBA, and average mobility was similar to the Lockdown Period, with a 49% gap from baseline. Last, by decomposing the intercity mobility matrix, the spatiotemporal mobility patterns of GBA and SBA were found to be very different. The three major mobility patterns in the GBA were daily, homecoming, and RTW, while the mobility patterns in the SBA were daily, weekday travel and weekend travel.

Although cell phone data can be a key tool for assessing the impact of COVID-19 interventions, data collecting mechanisms can differ between different regions and carries. These differences are often very difficult to balance. Therefore, it is important to understand the underling mechanisms of data collection and the associated limitations. In this research, the utilized SafeGraph dataset is derived based on POI visitation, which only collects user visitations to POI locations. The China Unicom dataset is stay-point based, which collects user visitations to all locations. Targeting to overcome this limitation, the POI dataset for GBA in 2020 was collected and a preliminary analysis was conducted by calculating the proportion of POI visitations in China Unicom dataset. Approximately 94.3% of the travel entries are POI-based. In addition, it is important to note that the mobility analysis for both GBA and SBA were conducted based on their own baseline (January 1 to January 7 for GBA and March 1 to March 7 for SBA) before their respective lockdowns. Because of the differences between GBA, SBA, and their analytical datasets, interpretation of the results should focus on their vertical changes instead of horizontal comparisons to avoid arriving at biased interpretations.

CRediT authorship contribution statement

Leiyang Zhong: Conceptualization, Methodology, Formal analysis, Software, Visualization, Writing – original draft. Ying Zhou: Conceptualization, Methodology, Funding acquisition. Song Gao: Resources. Zhaoyang Yu: Visualization, Software. Zhifeng Ma: Visualization, Software. Xiaoming Li: Conceptualization, Funding acquisition. Yang Yue: Funding acquisition. Jizhe Xia: Conceptualization, Methodology, Writing – review & editing, Supervision, Project administration.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

The following are the Supplementary data to this article: