1 Introduction

Estimating the total long-term economic impacts of disasters with economic models such as computable general equilibrium (CGE) provides the ability to analyze behavioral responses, economic resilience activities, and policy measures (Rose et al. 2017; Prager et al. 2018; Walmsley et al. 2021a). However, it is critical that these responses are captured with accurate data. The collection and interpretation of Earth observation (EO), such as that obtained from satellites, can augment publicly available economic data. Earth observation data are frequently used following major disasters to provide situational awareness of the extent of impact, which can often be detected by a hazard footprint (that is, extent of flooding or wind damage) (Huyck et al. 2014). In many cases, direct damage can be detected through automated routines or manual interpretation, which can form the basis of crowd-sourced impact assessment. Earth observation data are also used to infer the distribution of built assets, or to directly identify and quantify the vulnerability of specific buildings or assets, including critical infrastructure (Huyck et al. 2022). Expansion of the use of EO data into a total economic consequence modeling environment has the potential to increase the accuracy of assessments under conditions where default parameters or data prove to be inadequate.

We investigated whether crowd-sourced interpretations of very high-resolution optical satellite images from Maxar can be used to illuminate key economic behavioral responses for CGE modeling of the effects of COVID-19 in Los Angeles (LA) County, including: (1) Compliance with stay-at-home mandates; (2) Avoidance of economic activity based on a perceived threat or danger; and (3) Demonstration of resilient activities, such as recapturing lost production and production input substitution. We advanced the state-of-the-art by using EO data in the CGE context to improve accuracy and provide higher resolution for space and time than in prior studies. After a literature review identifying key contributions, we outline the methods, whereby we: (1) Gathered EO data of sector-based behavioral changes during 2020 and the first quarter of 2021, as compared to 2019; (2) Translated sector-based changes to CGE inputs; (3) Used CGE modeling to simulate key causal factors; and (4) Conducted dynamic adjustment calculations. We then present results, discussion of implications for research and policy, and conclusions.

2 Literature Review

There has been limited use of EO in economic modeling in the literature. Moran et al. (2020) reviewed new approaches connecting EO to economic decisions, focusing on global supply chain models and argued that EO-derived spatial datasets can help overcome their spatial and commodity resolution limitations. Escobar et al. (2020) used EO to detect land use changes and quantify greenhouse gas emissions embedded in agricultural production through a hybrid approach integrating life cycle assessment principles with the spatially explicit information on production to consumption systems (SEI-PCS) model of trade flows. The SEI-PCS model (Godar et al. 2015) allows for fine-scale sub-national assessments of the origin of, and socioenvironmental impacts embedded in, traded commodities. Diniz and Ferreira Filho (2015) analyzed the economic impacts of the Brazilian Forest Code using an interregional CGE model (TERM-BR), combining satellite imagery data with economic information from the Brazilian Agricultural Census. Tanaka et al. (2019) used remote-sensing (RS) data to estimate cereal yields one or two months before harvest with a world trade general equilibrium model to quantify the economic benefits of this RS forecasting triggered response.

In the closest example to our approach, Liu et al. (2017) used land use data derived from remote sensing imagery to improve efficiency of agricultural water use with a static CGE model. In this manner, EO data can contribute to CGE modeling of disasters by providing consistent indicators across a range of behavioral and economic activities before, during, and after the disruptive event. Earth observation data can also capture event dynamics in terms of key economic consequence factors identified in prior literature (Rose et al. 2017; Prager et al. 2018; Walmsley et al. 2021a, 2021b). This includes behavioral responses in the immediate aftermath of events, compliance with government policy (for example, public health orders, economic shutdowns, or shelter-in-place mandates), voluntary avoidance and aversion behavior (Rose 2022), and economic resilience strategies such as substitution, pent-up demand spending, and business recapture (Rose 2017; Wei et al. 2020; Dormady et al. 2022).

Identifying appropriate indicators to use in CGE simulations depends on the disaster type and economic consequence factor (Rose et al. 2017; Prager et al. 2018), and often numerous different event-specific data sources are combined with general (that is, non-event specific) findings from the literature. Recent studies have used CGE to examine the economic consequences of COVID-19, though primarily at the national level in the United States (Dixon et al. 2020; McKibben and Roshen 2021; Walmsley et al. 2021a, 2021b; Walmsley et al. 2023), and none focusing on LA County or using EO data. Earth observation data appear to be unique in providing consistently measured indicators across each of the economic consequence factors and across the timespan of disaster events. Most other studies used survey or government statistics, having the disadvantage of not being real-time data, which are more valuable for policy interdiction in terms of public health or the promotion of the economic recovery.

3 Methods

Our analytical approach includes four elements, as outlined in the following sections. Section 3.1 summarizes each step of the EO data collection process. Section 3.2 outlines the translation of EO sector-based changes to CGE inputs, which is a key contribution of this study. Section 3.3 provides a brief overview of CGE modeling of disasters. Section 3.4 proposes a new dynamic adjustment calculation approach to account for the length of the COVID-19 event.

3.1 Earth Observation Data Indicators

We used EO of automobile activity in parking lots and on adjacent streets in very high-resolution satellite imagery to provide sector-based behavioral changes during 2020 and the first quarter of 2021, as compared to 2019. The EO data collection involves seven steps. For Step 1, sector sampling, the team identified sectors of focus based on economic importance to the LA region and the added value of EO data for economic modeling (see Table 1). We were guided by three economic factors: compliance, avoidance, and resilience. Compliance refers to whether non-essential businesses are opened against public health orders. “Essential businesses” in LA County were defined by the “Safer at Home Order” in March 2020 (County of LA Department of Public Health 2020); around 34% of jobs and 16% of output were non-essential.

Table 1 Selected sectors for Step 1 of Earth observation data collection and number of parcels for Step 4
Full size table

Avoidance refers to a departure from normal or routine conduct due to fear (Giesecke et al. 2012; Prager et al. 2017; Rose 2022) that originates from the social amplification of risk, where people refrain from certain activities and minimize contact with others (for example, using alternatives to public transportation, reductions in tourism or business travels, staying home from work, and keeping children home from school). Here we identify which essential businesses closed or reduced operation due to avoidance, and after reopening, when and how much are people returning to prior behaviors.

Resilience refers to tactics that businesses are able to implement to offset lockdown-related losses with other activities or alternative practices (Rose 2007, 2017), and in the COVID-19 context include telework, business recapture (increased operations to offset losses), substitutions (increased spending on home improvement, greater use of e-commerce, and regional vacations), and pent-up demand. Pent-up demand is generated from the inability to spend during the mandatory closures and stay-at-home orders (Walmsley et al. 2021a), and includes increased car sales, air travel, spending on restaurants, apparel retail, and hotels after long closure periods.

For Step 2—identify and geolocate economic activities—we used the LA County Assessor land parcel boundaries geodatabase (for the random selection of samples), the Esri Business Analyst, and Google Earth. Additional data and analysis were required for the aerospace and textile industries. Esri Business Analyst data on more than 13 million U.S. businesses—including the business name, location, and industry classification code—were used to enhance the sectoral resolution of the LA County Assessor parcels. Google Earth was used to verify the sector labeling of each sample and in some cases to augment the number of samples (for example, for aerospace, textile, and distribution centers).

For Step 3, we screened samples based on a set of rules established to deduce demand from parking using Google Earth, and discarded if the sector was mislabeled or not clearly represented. These included: Medical and dental sector had to include only activities that were not related to COVID-19 in order to analyze avoidance behavior; Samples for manufacturing of textiles sometimes included wholesalers that had to be discarded; Businesses with shared parking lots were often discarded (for example, dental practices next to supermarkets or pharmacies), especially those that were close to an essential business that could contaminate results; Car dealerships required the identification of the clients’ parking space; Manufacturing and refineries required identifying workers’ parking areas; Restaurants had their own parking spaces in some areas and relied on street parking in other more central areas; Businesses with covered or partially covered parking lots were necessarily discarded.

For Step 4, the process identified a total of 341 selected parcels by sector (see last column of Table 1). The uneven sectoral distribution reflects the difficulty of locating certain sectors (for example, textile manufacturing, distribution centers, aerospace) and the parking characteristics of others (for example, shopping malls usually with covered parking space, dealerships where client parking is hard to identify, supermarkets with shared parking). Once the screening was completed, polygons corresponding to each parcel were edited to include only the parking areas that would be analyzed in the next steps.

For Step 5, we first selected multiple EO images to cover the group of 341 sector samples in the LA area for a series of dates spanning from 2019 (baseline) through 2021 (COVID-19 pandemic), using only Tuesdays, Wednesdays, and Thursdays to reduce traffic variability seen on weekends. This filtering resulted in 61 images being retrieved from Maxar’s catalog—which was limited for the year 2020, such that almost half of the images selected corresponded to 2019—yielding a total of 2318 sample images for analysis. The Maxar images were produced by three of Maxar’s sensors; GeoEye-1, WorldView-2, and WorldView-3. GeoEye-1 offers a panchromatic resolution of 0.41 m (16 in) and multispectral imagery at 1.65 m (5.4 ft) at nadir, covering expansive 15.2 km (9.4 mi) swaths. WorldView-2 provides panchromatic imagery with a resolution of 0.46 m (18 in) and eight-band multispectral imagery at a resolution of 1.84 m (72 in). WorldView-3 provides panchromatic imagery with 0.31 m (12 in) resolution and an eight-band multispectral imagery with a resolution of 1.24 m (4 ft 1 in). In all cases, images were initially retrieved as separate panchromatic and multispectral raster datasets and were then pan-sharpened to produce color images at roughly 50 cm resolution to be used for analysis. All Maxar collection times were between 18:25 and 19:15 UTC (10:25 a.m.–11:15 a.m. PST).

In Step 6—crowdsourcing via Amazon Mechanical Turk—anonymous workers were given detailed instructions to count vehicles in parking lots and to indicate “NA” when images were unsuitable for counting. To increase anonymous crowdsourcing worker reliability an upper threshold of 100,000 ft2 was applied to samples, resulting in a final set of 1772 sample images for crowdsourcing analysis. A total of 740 sample images were available for 2019, 778 for 2020, and 254 for the first quarter of 2021. For each sample image, three unique observations were collected from distinct crowdsourcing workers, resulting in a total of 5316 unique observations collected. To comply with Maxar’s data protection requirements, the imagery was clipped, rotated, and obfuscated to be safely shared while preventing reconstruction of useful spatial information.

In Step 7, before EO indicator development, the data required checking for reasonableness. We discarded conflicting results between the three MTurk workers. There was a high dispersion in the answers the workers provided for samples with over 100 cars, and for that reason most were discarded. Results were aggregated by quarter, with the distinction of pre- and post-pandemic period for the first quarter of 2020, and normalized with respect to the year 2019. Aggregation by quarter allows to better compare the EO results with economic datasets, and to use these results in CGE modeling. Finer aggregations following the timeline of closures were not possible given the low availability of imagery during 2020.

3.2 Translation of Sector-Based Changes to Computable General Equilibrium Drivers

Computable general equilibrium modeling of the economic impact of COVID-19 on the LA County economyFootnote 1 requires the identification of economic consequence drivers. The drivers shown in Table 3 are discussed in the following subsections, following the categories used by Walmsley et al. (2021a, 2021b) and Walmsley et al. (2023).

As shown in Table 2, we analyze EO data for each sector, and then compare them with two other relevant economic data sources available at the LA County level and across numerous sectors. First, California Employment Development Division employment data cover most sectors; however, the data are only available quarterly and are an indirect measure of economic production, as hiring and firing might lag. During COVID-19, federal government Paycheck Protection Program loans to businesses may have further obscured economic activity. Compared to employment data, EO data have more variance (standard deviation of 0.435 compared to 0.109) and are higher in value overall, suggesting more economic activity than indicated by employment figures alone.

Table 2 Analysis summary, earth observation (EO) indicator contribution, and computable general equilibrium (CGE) modeling differences with base case
Full size table

Second, retail sales data from California Department of Tax and Fee Administration (CDTFA) provide accurate data on household spending across many retail and service sectors, but no insight on other key sectors such as manufacturing, warehousing, health, or education. This retail data, when indexed, also have a lower average standard deviation (0.09) than the EO data, suggesting that the EO data may include more “noise.” This is particularly relevant for retail sectors, as EO data show customers visiting businesses, but not how much they are purchasing.

Overall, EO data suggest more activity than sales at restaurants and supermarkets, both of which might be explained by increased delivery and takeout levels. However, EO data suggest less activity than sales at hardware/building materials stores and car lots, both of which might be explained by costumers being more selective in their visits and purchases. For textile manufacturing, (non-COVID) healthcare, and aerospace manufacturing, EO data suggest lower economic production during the lockdown and COVID-19 surge periods than national data. Industry-specific data are available for warehousing and storage, though they are an indirect indicator of port container activity. In this case, EO data provide additional insights as to the dynamics of the pandemic, with greater activity during early periods when demand levels and inventories were high.

3.2.1 Mandatory Closures

COVID-19 public health responses in LA County began on 15 March 2020, with the “Safer at Home” order and were also covered by state-level orders. As shown in Fig. 1, county and state public health orders for our sectors of focus continued to evolve through the COVID-19 pandemic. Based on California Safer at Home and Blueprint for a Safer Economy mandates, we identified three sector groups:

  1. (1)

    Essential businesses (no modeling of closures). These remained open and operational, possibly with restrictions, for example, distancing and masking requirements.

  2. (2)

    Non-essential businesses that rely on physical presence (modeling of proportional closures). These businesses may have been closed for periods (for example, performing arts, movie theaters) or open with significant and changing restrictions (for example, restaurants, hair salons, gyms).

  3. (3)

    Non-essential businesses with high telework potential (modeling of telework labor impacts). These businesses, such as business services, education, religious and civic organizations, and government, were able to continue operating, but required to work remotely using telework.

Fig. 1
figure 1

Duration of closures due to COVID-19 for selected economic sectors in Los Angeles County, California from 15 March 2020 to 19 April 2021

Full size image

We focus on closures to non-essential sectors—such as textile manufacturing, retail trade, business services, education, arts and live entertainment, hotels, restaurants and bars, personal services, religious and civic organizations, and government—as these sectors rely on customer physical attendance to operate. Sector impacts are modeled in Simulations 1a (base case, using national data to estimate closures), 1b (adjusted for compliance, using national data to account for compliance), and 1c (adjusted for compliance using EO data to account for compliance) using a productivity parameter adjustment in the CGE model to allow sector production to be reduced indirectly and iteratively to direct impact shock levels.Footnote 2

Closure estimations are calculated according to the share of days in quarter that non-essential business was ordered closed (for example, in 2020 Q1, non-essential businesses were ordered closed for 16 days, or 18% of the quarter), the subsector share of LA County sector (for example, furniture and home furnishing, a non-essential retail subsector, is 3.9% of LA County aggregate retail sector), adjustments for modified openings (for example, in the California Tier system, businesses such as restaurants, nail salons, hair salons, retailers, and shopping malls could operate indoors at 25% capacity while cardrooms, gyms, museums, and wineries, were only able to open outdoors), and sector-specific factors (for example, restaurant closures were adjusted for the proportion of dine-in and off-site—delivery and pick-up—sales prior to the pandemic as well as the ability to dine outdoors). Significant increases in off-site and outdoor dining are examples of adaptive resilience substitution tactics.

Non-essential businesses with a high telework potential—business services, education, religious and civic organizations, and government sectors—faced transition costs yet reduced operating costs such as office space rent and travel while also increasing productivity (Prager et al. 2022). In Simulation 6a, telework direct impacts represent only additional telework during COVID-19. Our mandatory closures category does not include the counterfactual of companies closing if telework were not possible. Instead, we simulate telework as a resilience factor using labor factor increases in the CGE model.

3.2.2 Avoidance Behaviors

Avoidance behaviors for Simulations 2a and 2b (workplace and sector avoidance) are identified using the approach outlined in Walmsley et al. (2021a, 2021b):

  • Staying home from work is modeled as a − 4.58% labor force productivity impact.

  • Keeping children from school (impact on education sector) is modeled as a − 1.59% reduced demand in the education.

  • Reduction of in-person school attendance (caregiver impact) is modeled as a − 2.71% labor force impact.

  • Avoiding medical professionals is modeled as − 29.16% reduced demand for healthcare (compared to − 15.08% in the EO data).

  • Reducing shopping is modeled as − 5.85% reduced demand in wholesale and retail trade (EO observations were not reliable for shopping malls).

  • Avoiding local leisure activities is modeled as − 21.17% reduced demand for entertainment (EO observations suggest this to be − 37.9% in LA County).

  • Avoiding dining out is modeled as − 26.5% reduced demand in hospitality (EO observations suggest this to be − 10.1% for LA County).

  • Avoiding public transportation is modeled as − 4.21% reduced demand in public transit (Google Mobility suggests this is − 43.13% for LA County).

  • Cancelling air travel is modeled as − 57.48% reduced demand in air transport (LAX passenger data suggests this is − 46.82% for LA County).

Most of these proportional impacts are applied to the LA County CGE (LACGE) model (see Rose et al. 2007) for the base case simulation (Simulation 2a); however, as the sectoring scheme is different between the model used in that study (Walmsley et al. 2021a)—Global Trade Analysis Project (GTAP) (Hertel 1997)—and the LACGE model (LA County only), some sectoral adjustments are made. Simulation 2b uses EO data to adjust the reduction in household spending for healthcare, entertainment, and restaurants, though EO data collected for retail were found to be unreliable. Data on public transit (from Google Mobility) and air travel (LAX) passenger journeys were used to provide alternative values for these sectors.

3.2.3 Labor and Healthcare

Lost labor productivity from COVID-19 illness and caring for sick family members is calculated for LA County using the approach outlined in Walmsley et al. (2021a) and COVID-19 case, hospitalization, and fatality data from the LA Times (2022), BLS labor force participation rates (U.S. Bureau of Labor Statistics 2022), and Census data for LA County (U.S. Bureau of the Census 2022). COVID-19 case, hospitalization, and fatality data are multiplied by per patient lost productivity factors provided in Walmsley et al. (2021a). Increased demand for healthcare by quarter is estimated using the approach outlined in Walmsley et al. (2021a). Expenses by treatment category and age group are multiplied by the hospital patient (non-ICU and ICU separately) and outpatient numbers.

3.2.4 Resilience (Pent-up Demand)

For pent-up demand, we aimed to capture consumer spending constrained in one period—due to lockdown measures, avoidance behaviors, and so on—and in effect reallocated to later periods once pandemic conditions have abated. We identify the lowest quarter of consumer spending using best available sectoral data. For subsequent quarters we calculate the difference between expected business operation levels and the observed or actual spending unique to each sector, using LA County data where possible and national data where appropriate. We use quarterly California Department of Tax and Fee Administration data (CDTFA 2022) to calculate spending changes on retail goods (automobiles, apparel, general merchandise) as Angelenos may have different spending habits to other U.S. regions given unique public health measures and higher-than-average unemployment rates locally.

As LA attracts tourists from across the United States and worldwide, air travel, hotels, and restaurants are all likely particularly sensitive to mask mandates and international border closures. Conversely, it is possible that Californians increased local tourism to avoid air travel, offsetting negative impacts. Data on LA air passengers (LAWA 2022), seated diners at restaurants (OpenTable 2022), and hotel stays (Visit California 2022) are used to more accurately estimate pent-up demand for these sectors.

Restaurants increased seated diners from 3.5% of pre-pandemic levels for 2020 Q2 to 20% for 2020 Q3. However, this figure remained similar until 2021 Q1, with seated diners increasing to 58% in 2021 Q2, following the lifting of restrictions. For seated diners, it seems that pent-up demand may have shifted until after restrictions were lifted or substituted towards delivery and pick-up spending. Earth observation data align with national industry data suggesting that overall restaurant demand increased earlier than seated dining, suggesting restaurants increased delivery and pick-up during COVID-19. This highlights the importance of EO data in observing resilience measures and for economic consequence modeling of resilience tactics during disasters.

3.2.5 Resilience (Production Input Substitution)

Substitution relationships are key to CGE models, capturing inherent ability to shift between groups of inputs into production in response to price changes or external shocks. Two additional substitution effects are focused on here. First, telework enabled many non-essential businesses to move away from offices without experiencing productivity losses. To model this effect, data from Walmsley et al. (2021a) on additional COVID-19 telework rates is adjusted to account for full time equivalence (FTE) and productivity changes realized from telework. These differences are then applied as a labor force increase impact to offset reductions elsewhere.

Second, e-commerce increased significantly and above trend during COVID-19, from 10.6% in 2019 to 14.6% in 2020. While some substitution to online shopping was within-sector and is built into CGE modeling, some switched from service and experiential consumption (such as concerts and sporting events) to in-home consumption including home repairs, electronics, and home goods. Total consumption reduction for these other sectors is added to retail as a substitution effect, up to the USD 7.2 billion estimated size of LA County e-commerce increase in 2020. This may underestimate impacts as local inputs, for example, labor inputs, are much larger for storefronts than e-commerce.

3.2.6 Resilience (Production Recapture)

While most manufacturing was classified as essential under the California orders, some businesses were shuttered early in the pandemic. However, there is not strong evidence of production recapture in LA County employment data, EO data, or national production data for textiles, chemical products, and petroleum refineries so we do not include production recapture factors in manufacturing sectors. There is evidence to suggest production recapture took place for trade and warehousing, during late 2020 and early 2021. Twenty-foot equivalent unit (TEU) container data for LA and Long Beach ports and national level warehousing and storage data show an increase above prior year levels from 2020 Q3 onwards. We use these data to estimate production recapture factors for the wholesale trade sector of the LACGE model. There is also evidence that medical and dental employment in LA County increased during the study period. National level data for medical sectors suggests an increase after a quarter of reduction, however not above prior year levels. As such, these sectors are not included in the base case simulations.

3.3 Computable General Equilibrium Modeling

Computable general equilibrium analysis models the behavioral response of producers and consumers to price changes, new regulations, and external shocks in the context of markets, within the constraints of labor, capital, and natural resources (Rose 1995). We use a recently updated version of the LACGE model, which is a regional derivative of the USCGE model (Rose et al. 2007, 2017; Prager et al. 2018) based on IMPLAN data on the LA County economy. The LACGE model uses non-linear optimization algorithms to simulate economic activity across all economic sectors, households, government institutions, and trade with the rest of the United States and rest of the world. Producer decisions regarding input utilization are represented through a series of nested constant elasticity of substitution functions, whereby inputs are combined in a hierarchal process beginning with separate capital, energy, labor, and materials components and then aggregated pairwise. Elasticities in Constant Elasticity of Substitution functions can differ across nests, and they range from perfect substitution to no substitution. Values for substitution elasticities are obtained from the literature and tested for reliability. Labor and capital income payments from producing sectors are allocated to the nine household income brackets, and subject to appropriate taxation and depreciation; transfers between institutions are also represented. Household consumption is divided across aggregate commodity groups through a linear expenditure system, a series of demand functions varying with respect to income level. Household consumption incorporates household demands, price changes, and household substitution elasticities. Government consumption is represented with a fixed-coefficient expenditure function. Household and government savings are fixed proportions of disposable income (that is, income following adjustments for taxes and transfers) and are balanced in their respective equations by savings from foreign sources. Each of these institutions also borrows capital. Investments are financed by net institutional savings plus depreciation charges and retained earnings.

Computable general equilibrium models simulate major economic changes, or shocks, and their subsequent interactive effects, accounting for price changes and income and input substitution elasticity effects. Computable general equilibrium models are particularly useful for policy analysis when empirical data are limited, such as for modeling the economic impacts of disasters and the implications of alternative recovery paths.

We simulate causal factors in terms of mandatory closures, avoidance, labor supply, healthcare, resilience in terms of pent-up demand, substitution, and recapture, using a range of modeling approaches. The base equilibrium state in the LACGE model is “shocked” to a level reflecting empirically observed conditions (such as EO data). The new equilibrium state is checked for reasonableness across different model outcomes, for example, sectoral output and prices, labor and capital usage, and household consumption, before macro-level output and employment changes are identified. The final column in Table 2 provides the contribution of EO data with respect to CGE modeling approaches for each sector.

3.4 Dynamic Adjustment

Given the length of COVID-19, event dynamics are important to consider. As highlighted by Cole (1988), Okuyama et al. (2004), and others regarding the dynamics of disaster impacts, economic impact modeling often assumes “production simultaneity”, that is, that production activity and changes to it occur at the same time (Avelino and Hewings 2019); however, that assumption makes little sense in a long event such as COVID-19. Focusing on our CGE driver data, early in the pandemic period EO data (as well as some quarterly economic indicators) are likely to capture a purer COVID-related direct impact. The sudden shock of mandatory closures and other policy changes are likely to have been the primary and dominant factors influencing production decisions in this early period since it takes time for the general equilibrium effects to flow through the economy. As the pandemic developed, production and consumption decisions that EO data and other economic indicators represent are likely to increasingly manifest themselves as indirect and induced effects, such that later observations are likely to measure total effects.

To account for this, we utilized a post-CGE adjustment method to account for EO data capturing direct effects in the early period of a disaster event, and then an increasing amount of indirect (general equilibrium) effects as the event progresses. Adjustments are made only to simulations based on EO data, or to those where quarterly indicators face the same issue. For each quarter, we multiply economic indicator changes for each sector with variable CGE multiplier values.Footnote 3 The dynamic CGE multiplier value starts in the first period as equal to the total shock CGE multiplier value, and then increases or decrease to a value of 1 by the fourth period. For example, for Simulation 1b the total result CGE multiplier is 1.83 (total impact of − 0.86), so the CGE multiplier begins at this level in 2020 Q1 and declines linearly to a value of 1 in 2020 Q4 and 2021 Q1. EO data suggest negative shocks of − 0.31 in Q1, − 0.44 in Q2, and − 0.64 in Q4 of 2020, which are multiplied by the dynamic CGE multiplier values of 1.83, 1.55, and 1, respectively, to generate values of − 0.19, − 0.23, and − 0.22 to equal a total shock of − 0.64. This adjustment approach shifts the total impact from − 0.86 to − 0.64.

4 Results and Discussion

We examine base simulation results first. Mandatory closures are modeled as constrained output using the productivity parameter in the production function for relevant sectors at the levels identified in Table 3: − 16.7% to retail trade, − 28.1% to entertainment, − 30.3% to hospitality, and − 47.7% to personal services. The mandatory closures base simulation (Simulation 1a) results in the largest negative economic impacts in this analysis (− 6.95%) and a less severe but still significant reduction in employment (− 2.39%). These mandatory closure impacts capture only the estimated closures to non-essential businesses that rely on physical presence of employees, as opposed to those that could take advantage of telework.

Table 3 Computable general equilibrium (CGE) modeling approach and results
Full size table

Non-essential businesses with high telework potential were instead modeled as a resilience (substitution) strategy reflecting an increase to labor supply from telework (Simulation 6a). This captures additional telework in non-essential sectors such as information, finance and insurance, professional services, management and administrative services, education, and government during the pandemic (adjusted for full-time equivalence and productivity changes). This simulation suggests a substantial impact in absolute terms, of +10.7% to output, and +18.80% to employment. Telework appears to have helped mitigate substantial economic costs and job losses during the pandemic.

Of the substantial workplace avoidance across the US during the COVID-19 pandemic, most was attributed to two behaviors: avoiding exposure to the virus and caring for sick family members (Walmsley et al. 2021a). The base simulation representing workplace avoidance (Simulation 2a) is the second largest negative economic impact, resulting in a − 4.08% output change and − 7.47% employment impact. Per Simulation 2ci, avoidance also relates to in-person school attendance and household spending on retail, leisure, dining, public transit, and air travel and uses direct impacts values from Walmsley et al. (2021a). Education avoidance (Simulation 2b) has a smaller economic impact (− 0.20% change to output) than the combined household spending impacts (− 0.29% change to output).

Simulations 3 and 4 account for labor supply contraction from deaths and illnesses in LA County, and for increased healthcare spending. Labor supply impacts from deaths and illnesses are notable, accounting for a − 1.04% change in output and a − 1.93% change in employment. Like other labor supply shocks, the labor impact is larger proportionally than that to output, due to CGE labor-capital substitution effects reflecting sector ability to switch to more capital-intensive operations. Conversely production or household spending-related shocks result in larger proportional impacts to output than employment.

In addition to Simulation 6a, resilience is modeled with production recapture substitutions (Simulations 7a and 7b). Before dynamic adjustment, pent-up demand using this data increases output by 0.07% and employment by 0.04%, a relatively small offset. However, as the CGE multiplier here is less than 1, and pent-up demand spending is the largest in the later stages of the study period, the adjustment increases results to 0.27% for output and 0.16% for employment.

Focusing on the EO data simulations, Simulation 1a uses averages of national and LA County employment indices while Simulation 1b uses EO data to capture the effect of closures to the textile manufacturing sector. Textiles was one of the few manufacturing sectors that appeared to be non-exempt from the California Stay at Home order. Pre-adjustment, lower EO data direct impacts result in lower total impacts compared with national data. However, as the CGE multiplier is less than 1, the dynamic CGE multiplier values increase the total result up to -0.04% for output despite the shock values occurring in 2020 Q1, Q2, and Q4.

Simulations 2ci, 2cii, and 2ciii compare different measures of household spending avoidance during the pandemic. Household spending changes for numerous sectors—retail, entertainment, restaurants, public transit, and air travel—were calculated with EO data, Google Mobility data, and LAX passenger data, and compared against data from Walmsley et al. (2021a). Earth observation data suggest lower avoidance of non-COVID-related healthcare and dining out during the pandemic than national data. Conversely, EO indicators of leisure activity avoidance is higher than national estimates. When combined with higher levels of public transit avoidance in LA County according to Google Mobility data (Google 2022), the overall economic impacts of household spending avoidance are more than double the base simulation. For Simulations 2cii and 2ciii, dynamic adjustment calculations are based on CGE multiplier values less than 1, and hence increase the total impact up to − 0.83% and − 0.5% for output, respectively.

For pent-up demand, EO data on numerous sectors—retail (car sales, apparel, general merchandise, hotels, wellness and fitness), restaurants, live experiences, real estate, air travel, supermarkets, and non-COVID (ambulatory) healthcare—are compared against the best available data. Earth observation data are unavailable for real estate and air travel, so base values are used to allow for comparison. Earth observation data suggest that there is much larger pent-up demand for services, groceries, and non-COVID healthcare. Altogether, EO data simulations have double the impact to output and 50% more impact to employment than the base simulations.

Analysis of production recapture (see Sect. 3.2.6) suggests that it was only consistently empirically observed for warehousing and storage. Other commonly used approaches focus on recapture of output lost during the shock period (with diminishing effect after the first three months); however, our approach compares EO data with the best available data. The base simulation (7a) of production recapture for warehousing and storage increases output and employment by 0.02%, which is adjusted up to 0.04% for output and 0.03% for employment. In comparison, Simulation 7b using EO data suggests a much larger positive impact to output (0.09%) and employment (0.08%), which is adjusted up to 0.11% for output and 0.09% for employment.

5 Conclusion

This study shows conclusively that EO data provide additional insights when modeling the economic impacts of disasters and other extreme events, especially when using CGE models. Earth observation data do so consistently, and with greater resolution, across space, time, and industry sectors. Improvements in terms of accuracy and additional resolution vary by economic sector. For manufacturing sectors, these data improve the accuracy of estimates by providing quarterly regional and sub-regional data that are otherwise only available at the national level. For warehousing and other supply chain-related sectors, EO data similarly provide a level of detail not attainable in publicly available quarterly regional data. For retail sectors, these data could be compared against sales tax receipts to provide both verification and further insights. There are constraints in the availability of EO data, however, even for a highly populated county like Los Angeles. Very high-resolution satellite imagery often requires further processing due to privacy and security concerns. Image density is insufficient to have a historic record. Future studies should consider tasking imagery in advance.

Comparing CGE simulations of EO data with those for the best available alternative data highlights the need for more widespread monitoring of economic activity and compliance with public health orders and policy directives during pandemic and disaster events. For example, EO data on the textile manufacturing sector suggests that the negative impacts were lower than for other data, while negative impacts from household spending avoidance related to healthcare were lower also. In contrast, EO data indicated larger negative impacts to leisure and retail spending than the alternatives. Simulations of resilience factors using EO data also estimated much larger impacts, though in the positive direction. These differences highlight the important contribution that remote sensing can make to our understanding of economic-related behaviors during disaster periods. The results further highlight the opportunities that EO can bring to the economic modeling of disaster impacts and to economic modeling in general.

We also highlight the importance of dynamics in a longer-term disaster event such as COVID-19. Earth observation and other economic indicators might be used effectively to capture direct impacts of disasters in the immediate aftermath of the event. Yet when disasters have longer impact periods, the general equilibrium effects must be accounted for, and hence EO and other economic indicators are likely to increasingly reflect total impacts rather than direct impact. Our dynamic adjustment approach accounts for this function and suggests that, when CGE multiplier effects are less than 1, the total impact will increase, and vice versa for CGE multiplier effects greater than 1. Our approach also accounts for the periods in which impacts are experienced; this further increases the impact of causal factors such as pent-up demand that is experienced further away in time from the initial shock.

Future research using EO data to capture the economic consequences of disasters using CGE modeling would be useful. The outcomes of this investigation underscore the need for a comprehensive exploration of the opportunities to integrate EO data and CGE models, especially in other natural and human-caused disaster contexts. This research shows the potential for EO data usage, while also highlighting the need for sensitivity to disaster- and sector-specific conditions. Increasing the spatial and temporal resolution of EO data, or alternatively the geographical scope, could also provide additional insights into the dynamics and geographical aspects of disaster economics. Its utility extends beyond merely characterizing the economic aftermath of disasters, encompassing a predictive capacity to anticipate potential consequences of likely natural hazard-induced disasters under both current and future climate conditions. Further research has the potential to introduce innovation to CGE modeling by strategically locating economic data with EO, thereby enhancing the accuracy of predictive analytics and fostering well-founded policy development.