Modeling housing inventory changes over space and time for natural hazards applications builds on three bodies of literature: landscape change models, housing economic models, and natural hazards exposure models.
2.1 Changing landscape models
Methods for modeling changing landscapes across large areas are rooted in the land use–land cover (LULC) change literature (Daniel et al. 2016; Sleeter et al. 2017). Many LULC models are used to predict urbanization changes over long periods of time, usually at decadal time intervals. The units of analysis for LULC models are typically at 1 km2 or less and can span city to national scales. The structures of LULC models vary depending on the application, including logistic regression-type models that predict urban versus non-urban areas over time (Song et al. 2018), classification-type models that identify the LULC category (urban, forest, agriculture, etc.) per area unit over time (Sleeter et al. 2017), and regression-type models that predict quantifiable changes per area unit such as the percent or urban land in an area unit (Gao and O’Neill 2019). Overall, LULC models for large spatial scales tend to utilize linear regression, cellular automata (CA) methods, machine learning methods, or a combination thereof (National Research Council 2014; Cremen et al. 2022). Aburas et al. (2019), Briassoulis (2019), Musa et al. (2017), and Verburg et al. (2004) provide reviews of different CA and machine learning methods for LULC modeling, as well as commonly used modeling parameters. Cellular automata models are usually set up as classification-type problems and develop a set of transition rules and transition probabilities where each area unit (often a grid cell) can change states based on neighboring characteristics. Machine learning models can be more flexible and designed as logistic, classification, or regression problems. Satellite imagery has been used with convolutional neural networks (CNNs) to classify changes in land cover types per pixel at a given time, and with recurrent neural networks (RNNs) to estimate time series landscape changes (Aburas et al. 2019; Briassoulis 2019; Musa et al. 2017; Verburg et al. 2004).
2.2 Changing housing inventory models
Housing economics and real estate models are often developed using either local supply-and-demand land value models or computable general equilibrium (CGE) models for regional-scale estimates (Cho et al. 2005; Ustaoglu and Lavalle 2017; Ali et al. 2020), or agent-based or system dynamics models for local housing estimates (Wheaton 1999; Parker and Filatova 2008; Magliocca et al. 2011; Filatova 2015). However, these models rarely estimate the spatial location of new housing units at a subcounty scale over a multi-county region.
Davidson and Rivera (2003) is an exception, where the researchers utilized county-level population projections, headship rates, along with census-tract-level housing data to predict the number, location, and types of housing units per census tract for 15 counties in North and South Carolina over 5-year intervals between 2000 and 2020. The Spatially-Explicit Regional Growth Model (SERGoM) also used county-level population projections, headship rates, historic housing block and block group census data, land data, and road network data to project the spatial distribution of housing unit density over a 100 m2 grid space at decadal intervals for the entire conterminous US (CONUS). Census block data were transformed to a uniform grid space using simple dasymetric mapping techniques, which assumes a homogeneous distribution of housing units across the blocks. Future changes in housing unit densities per grid cell were calculated by (1) applying average growth rates between decades for 16 classification types formed from four housing density levels and four road density levels, and (2) estimating the travel time along major roads from a cell to an urban area. The combination of the previous decade’s growth rate and the urban travel time forms the allocation weights that were used to distribute county-level population estimates to housing densities per grid cell via county-specific headship rates in 2000 (Theobald 2005). The Integrated Climate and Land-Use Scenarios (ICLUS) developed by the US Environmental Protection Agency (EPA) extended the SERGoM methodology for predicting decadal changes in housing density by sourcing population projections under five different climate change scenarios to produce five different housing density scenarios in each decade between 2010 and 2100 (US EPA 2023). While these three housing projection models provide subcounty housing unit estimates, all rely on population forecasts, only provide a temporal resolution of five years or more, and only source input housing data at the census tract or block groups level, rather than specific coordinates per house.
2.3 Changing natural hazards exposure models
As a community develops, its exposure to potential harm from natural hazards increases. This phenomenon is known as the “expanding bullseye effect” when the footprint of a metropolitan region’s urban, suburban, and exurban area expands, thus increasing the “bullseye” natural hazard impact (Strader and Ashley 2015). This section reviews previous studies that combine forecasted changes in urban land cover, population, or housing with natural hazard risk models for areas within the US.
2.3.1 Projected urban change models and natural hazard risk
Song et al. (2018) and Rifat and Liu (2022) both utilized machine learning methods to develop urban change models under a set of different scenarios for small regions within Florida to estimate the expected change in flood risk in 2030 and 2035, respectively. Sleeter et al. (2017) used a cellular automata LULC modeling approach to estimate future changes in land cover in the US Pacific Northwest at annual increments between 2011 and 2061 to estimate changes in tsunami risk.
2.3.2 Projected population change models and natural hazard risk
Hauer et al. (2016) used a modified version of the Hammer et al. (2004) method to predict decadal changes in census block populations under five development scenarios across CONUS between 2010 and 2100 to evaluate the expected change in flood risk due to sea level rise. Wing et al. (2022) also projected population estimates to understand expected changes in flood risk across the CONUS, although it used ICLUS’ gridded population projection in 2050 rather than census block estimates.
2.3.3 Projected housing change models and natural hazard risk
Jain and Davidson (2007a; b) combined spatially-explicit housing projections with a hurricane loss model to assess the expected change in risk over time. These studies specifically predicted the changes in the number, location, type, and vulnerability of wood-framed buildings per census tract to evaluate expected changes in hurricane risk for selected North Carolina counties between 2000 and 2020. Freeman and Ashley (2017) also evaluated the change in hurricane risk via housing projections sourced from ICLUS for six major US coastal metropolitan statistical areas (MSA). Sanderson et al. (2022) utilized an agent-based approach to estimate annual, parcel-level housing inventory changes under 10 different land use planning scenarios in combination with seismic and tsunami loss models in the community of Seaside, Oregon, 2010 to 2040. Strader et al. (2015) used housing density projection outputs from ICLUS to evaluate the change in volcano risk in the Pacific Northwest at decadal intervals between 2010 and 2100. Finally, Mann et al. (2014) predicted the expected decadal changes in housing density across split block groups in California between 2000 and 2050 to evaluate wildfire risk under business-as-usual growth, urban growth, and rural growth scenarios using statistical regression modeling methods. Similar to SERGoM/ICLUS methods, this study extrapolated housing unit projections from county-level population projections supplied by the state government.
2.3.4 Summary
Estimating changes in natural hazard risk to a region’s housing inventory over space and time requires a method for accurately estimating the future location of new housing units in a relatively high spatial resolution (e.g., subcounty) and over a relatively high temporal resolution (e.g., annual) time steps. To clearly and realistically incorporate hazard impacts such as high wind speeds or large flood depths in natural hazard risk models, it is critical that the analysis be high resolution, both temporally and spatially. None of the aforementioned studies specifically model spatial changes in a region’s housing inventory in subcounty spatial increments and annual time increments. SERGoM-based models, such as those presented in Freeman and Ashley (2017) and Strader et al. (2015), provide housing unit density projections over a 100 meter grid space, however these projections are only available at decadal increments. Additionally, the obtained spatial housing data is manipulated from census blocks to a grid space using simple dasymetric mapping assumptions and future housing estimates were derived from population forecasts alongside expected headship rates, rather than housing forecasts directly.
This paper adds to the existing literature by developing a method that explicitly estimates the location of new housing units to subcounty area units over annual increments in the future across a large, multi-county region. The method utilizes coordinate-based housing data (rather than census-derived data) and is driven by county-level housing forecasts (rather than population forecasts).