Classifying Building Roof Damage Using High Resolution Imagery for Disaster Recovery

Abstract

Post-hurricane damage assessments are often costly and time-consuming. Remotely sensed data provides a complementary method of data collection that can be completed comparatively quickly and at relatively low cost. This study focuses on 15 Florida counties impacted by Hurricane Michael (2018), which had category 5 strength winds at landfall. The present study evaluates the ability of aerial imagery collected to cost-effectively measure blue tarps on buildings for disaster impact and recovery. A support vector machine model classified blue tarp, and parcels received a damage indicator based on the model’s prediction. The model had an overall accuracy of 85.3% with a sensitivity of 74% and a specificity of 96.7%. The model results indicated approximately 7% of all parcels (27 926 residential and 4431 commercial parcels) in the study area as having blue tarp present. The study results may benefit jurisdictions that lacked financial resources to conduct on-the-ground damage assessments.

Introduction

In the United States, hurricanes cause substantial economic damage and are responsible for nearly 7000 deaths since 1980 (NOAA 2023). In response to hurricanes and other disasters, the Federal Emergency Management Agency (FEMA) provides aid to qualified communities. FEMA’s Public Assistance (PA) Program provides grants to governments and nonprofits after disasters to aid community recovery (FEMA 2023). The Individual Assistance (IA) Program is meant to supplement individuals’ insurance, providing funds for basic needs and recovery costs (FEMA 2022). IA has been found to be one of the most vital programs for recovery in the immediate aftermath of a disaster (Lamba-Nieves and Santiago-Bartolomei 2022). In addition, FEMA and the Army Corps of Engineers may offer the Blue Roof Operation after a disaster. The goal of this collaboration is to install blue tarpaulin (tarp) at no-cost on roofs of eligible homes to temporarily prevent further damage of homes until permanent repairs can be completed (FEMA 2017). While they are meant to be short-term solutions, blue tarps have been observed for years after an event (Rowan and Kwiatkowski 2020). Although tarps will not be used for severely damaged buildings, the presence of blue tarps still gives insight into how much damage communities experienced and how they are recovering (Miller et al. 2013; Miura et al. 2020; Naito et al. 2020; Rathfon et al. 2012).

The present study has two main objectives. First, we seek to quantify the number of parcels with blue tarp present in counties designated by FEMA for individual and public assistance (Area A) versus just public assistance (Area B) after Hurricane Michael (October 2018). Unlike previous work, the current study seeks to classify damage over a large geographical area, including severely and relatively minimally affected counties. We investigated whether accurate results could be obtained across diverse environments. Blue tarps are an indicator of post-hurricane destruction that are also relatively spectrally distinguishable from other infrastructure (Rathfon et al. 2012; Miura et al. 2020). Second, while multiple short- and medium-term recovery programs are based on county-level damage assessments, few studies have actually quantified the number of buildings impacted by severe conditions within all disaster affected counties (Gurley and Masters 2011; Walker 2011; Tomiczek et al. 2014). We determined whether there are small pockets of damage to households outside of counties eligible for IA. The results of the study provide a finer scale to evaluate damage and quantify how many households may be missed when relying on county-level damage assessments.

High-resolution remotely sensed data is capable of identifying blue tarps over the entire affected area. Ground-based damage surveys are accurate and reliable but frequently focus on a specific municipality with a limited sample size (Gurley and Masters 2011). Furthermore, many communities lack the resources needed to even conduct such surveys (FEMA 2016). Alternatively, remotely sensed data collection can aid in disaster response (e.g., routing relief supplies) and recovery efforts. Such data can assess damage over a wide area relatively quickly and cost-effectively when compared to ground-based surveys (Jiang and Friedland 2016). For this study, we used high spatial resolution (<0.5 ft2) optical imagery to detect blue tarps as indicators of small areas of damage after Hurricane Michael. While high-resolution imagery provides superior ability to detect small areas of damage, it has drawbacks in terms of the data size and expertise needed to use the data (e.g., access and skill to use high-performance computing clusters) (Sun et al. 2020).

High-performance computing has contributed to a growing body of work using machine learning and deep learning methods to classify imagery and study disaster impacts. One such method includes the support vector machine (SVM) classifier. SVMs have been applied to assess disaster damage and recovery with comparable results to deep learning methods such as convolutional neural networks (CNNs) (Hasan et al. 2019). Using SVM, researchers have accurately identified collapsed buildings after disasters demonstrating the method’s suitability in urban and rural contexts (Moya et al. 2018; Sheykhmousa et al. 2019). CNNs have produced highly accurate results, as well detecting different levels of damage (none, minor, major, and destroyed) with improvement over an SVM in Bay County, Florida and portions of its neighboring counties after Hurricane Michael (Berezina and Liu 2022). However, this study chose an SVM over deep learning methods since it is relatively accurate with smaller training data and shallow data structures (Wang et al. 2021). We relied on publicly available optical imagery with only three bands to detect blue tarps after Hurricane Michael.

The results of the classification were used to determine whether blue tarps were located in counties ineligible for IA. In response to environmental disasters such as hurricanes, governors of affected states can request a Major Disaster Declaration based on a variety of factors, including damage severity and state funds available. If accepted, federal programs including IA, PA, and/or Hazard Mitigation Assistance may be available based on the request and preliminary damage assessments (FEMA 2022). Recent evidence has shown that certain vulnerable populations are more likely to be denied aid even if located in an eligible area (García 2022; Lamba-Nieves and Santiago-Bartolomei 2022). After Hurricane Maria struck Puerto Rico in 2017, large numbers of homeowners were denied aid for reasons such as being unable to prove homeownership and scheduling inspections (García 2022). In addition, areas with higher populations of foreign-born individuals were found to have lower damage estimates and therefore less aid (Grube et al. 2018). However, when looking at social vulnerability factors and damage at the county level, damage severity is still the largest predictor of IA aid (Emrich et al. 2022). This is expected and gives credibility to the way the program is operating, but county-level assessments still leave opportunity for households to be missed.

Methods

Study Area and Event

Hurricane Michael made landfall in Bay County, Florida on October 10, 2018. The storm contributed to 45 fatalities and over $25 billion in damages (NOAA 2023). Power outages spanned six US states and included 1.7 million customers (Woodward and Marcy 2018). The hurricane was the strongest on record to hit the Florida panhandle and, as a result, federal disasters were declared for counties across the area. This study covered 15 of those counties located in the Florida panhandle: Bay, Calhoun, Franklin, Gadsden, Gulf, Hamilton, Holmes, Jackson, Jefferson, Liberty, Madison, Taylor, Wakulla, Walton, and Washington. These counties generally had higher rates of poverty (12.2–29.9%) and populations without health insurance coverage (8.9–18.3%) than national rates (11.8% and 8.8%, respectively). In addition, these counties had lower rates of employment than the national average as well as lower average educational attainment (US Census Bureau 2018). For the purpose of quantifying damage at a building level, we relied on parcel data. Parcels are legally owned areas of land with a defined location, boundary, and a specific type of usage or development designation. In total, the data included 461 106 parcels located within the study area. Of this total, 101 596 of the parcels were commercial, with the remaining 359 510 parcels (78%) classified as residential. FEMA designated these 15 counties eligible for assistance following Hurricane Michael. The most severely impacted counties qualified for IA and PA from FEMA. These included Bay, Calhoun, Franklin, Gadsden, Gulf, Holmes, Jackson, Liberty, Taylor, Wakulla, and Washington counties. The remaining four counties qualified for PA alone (Figure 1) (FEMA 2018).

Figure 1.

Map of the Florida panhandle showing counties that were affected by Hurricane Michael. Federal Emergency Management Agency (FEMA) A indicates a county was eligible for both individual and public assistance. FEMA B counties were only eligible for public assistance.

Data Sources

The post-Hurricane Michael imagery used in this study were sourced from aerial photographs taken through the Florida Department of Revenue and stored in the Florida Department of Environmental Protection’s Land Boundary Information System (LABINS) (LABINS 2019). The images were captured between December 2018 and February 2019, and have 0.5-foot spatial resolution with red, green, and blue spectral bands. Each image tile covers one square mile and in total the data required 5 terabytes of storage.

Parcel Data

Parcel data were sourced from the University of Florida’s GeoPlan Center and contain the Florida Department of Revenue’s tax roll information from 2018 (FGDL 2019). In addition to information about the property (e.g., year the structure was built, value, use), these data contain spatial polygon information. When matched by location in a Geographic Information Systems software, these data can be used to determine how many households had blue tarps present within the study area. These parcel data are publicly available.

Image Classification

The study area includes a variety of landcover types. However, the class of interest only included blue tarps to indicate whether a structure was damaged or not. Therefore, the final classification result was obtained through two steps: first, a multi-class classification and second, a binary classification distinguishing blue tarps from all other classes. The multi-class schema contained seven training classes: Blue Tarp, Pool Water, Impervious Surface, Vegetation, Bare Soil, Roof, and Large Waterbody. These training classes are created by taking samples of pixels belonging to these different classes from different areas in the image. Researchers then created a signature file in ESRI ArcMap software using these training samples. We manually entered the mean values from the signature file to train the “Train Support Vector Machine (SVM) Classifier” in Python 3 using the scikit-learn package (Pedregosa et al. 2011). SVM classifiers distinguish classes by creating a linear hyperplane to separate the data. While SVM was originally used for binary classifications, the Support Vector Classification tool in scikit-learn implements a “one-versus-one” approach for multi-class problems. This approach reduces the problem to a collection of binary problems (Foody et al. 2006). SVM models are commonly applied to classification problems and perform relatively well compared to other deep learning methods even with limited training samples (Jozdani et al. 2019).

Batch processing tools were required to not only classify the aerial photographs (raster tiles), but also to convert the output into vector data for intersection with parcel information. A scripted batch code handled data classification and processing. The Rasterio package (GitHub 2019) read in the imagery as three-dimensional Numpy arrays (Harris et al. 2020). Each image was flattened to a two-dimensional array and run through the SVM classifier. The results were recoded to a binary classification (tarp or not tarp) and then converted to polygons using the Fiona package (GitHub 2019) to compare to the parcel information. We separately assessed the accuracy of the 1) multispectral classification and 2) multispectral classification with a spatial filter which excluded tarp polygons less than 200 square feet to reduce noise in the classification. Sensitivity testing filtering smaller polygons (50 and 100 square feet) did not reduce the noise while removing areas larger than 200 feet resulted in more false negatives and reduced the overall accuracy.

The accuracy assessments were based on a separate validation sample of 300 randomly positioned points (150 points per class). The points were placed within parcels with a minimum of 328 feet (100 meters) between points. Each point was manually checked for correct classification by looking for the presence or absence of visible blue tarps within a parcel in the aerial imagery. Researchers then calculated the overall and target class accuracies for the multispectral classification and multispectral classification with a spatial filter.

Results

The results of the classification accuracy assessment are reported in a confusion matrix (Table 1). The accuracy assessment was conducted for both the spatially filtered and unfiltered results. Before the application of the spatial filter, the overall accuracy was 78%. The final filtered results indicated an overall accuracy of 85.3% with a kappa coefficient of approximately 0.71, indicating an improvement over the result that was not filtered. The user’s accuracies of the blue tarp class and non-tarp class in the filtered result were 95.7% and 78.8%, and producer’s accuracies were 74% and 96.7%, respectively. Pools and roofs, which were highly reflective, were more commonly confused with blue tarp than impervious surfaces, vegetation, or darker roofs. The spatial filter by polygon area improved the specificity of the tarp classification from 82% to 96.7%.

Table 1.

Two class accuracy assessment of 300 independent and randomly distributed points. The spatially filtered results (bottom) show improvement in the overall accuracy and specificity over the unfiltered results (top).

Unfiltered
Class	Tarp	Other	Total	User’s Accuracy:
Tarp	111	27	138	80.4%
Other	39	123	162	75.9%
Total	150	150	300
Producer’s Accuracy:	74%	82%		Overall Accuracy: 78%
Spatially Filtered
Class	Tarp	Other	Total	User’s Accuracy:
Tarp	111	5	116	95.7%
Other	39	145	184	78.8%
Total	150	150	300
Producer’s Accuracy:	74%	96.7%		Overall Accuracy: 85.3%

The presence of any blue tarps indicated whether a parcel had been damaged by the hurricane and received a damage designation in the classification (Figure 2). Within the study area, blue tarps were located in approximately 7% of all parcels (27 926 residential and 4431 commercial). Examining just residential parcels by county, results showed Calhoun County had the largest proportion of all parcels with blue tarps at approximately 22.6%, followed by Bay County at 20.5% and Jackson and Gulf Counties at 16.9% and 13.9%, respectively (Table 2). In general, the counties in Area B including Hamilton, Jefferson, Madison, and Walton had relatively lower percentages of parcels (0.3–2.9%) with tarps, which is in agreement with the FEMA aid classification. These counties were all located farther from the hurricane track (Figure 3).

Figure 2.

The unaltered aerial photograph is shown (top left) followed by the results of the classification overlayed in orange (top right). The model classified tarps in all counties which had imagery available and were disaster designated by Federal Emergency Management Agency (FEMA) (bottom).

Table 2.

The proportion of residential parcels in each county where the model detected blue tarps as a percentage of all residential parcels within that county.

County	Parcels with Tarps (%)
Area A
Bay	20.5
Calhoun	22.6
Franklin	3.5
Gadsden	7.0
Gulf	13.9
Holmes	4.8
Jackson	16.9
Liberty	9.3
Taylor	1.2
Wakulla	1.1
Washington	7.3
Area B
Hamilton	0.7
Jefferson	0.9
Madison	0.8
Walton	3.5

Figure 3.

The map shows the percentage of residential parcels with tarps by county with high proportions located in counties nearer to the hurricane track.

Discussion

Main Findings

The results of this study demonstrated that we could determine individual building damage across a large disaster-affected area at relatively low cost and time. Other studies using various methods, including CNNs and combinations of remote sensing and survey data, have resulted in higher accuracies. However, these studies generally focused on a smaller area or on just one or more urban centers (Berezina and Liu 2022; Gurley and Masters 2011; Miura et al. 2020; Naito et al. 2020). However, the present study covered a much wider geographical area with more variability in environment (urban versus rural, coastal versus inland) and included photographs taken on different days, which may result in differences in reflectance of blue tarps. Despite these challenges, the SVM model used still resulted in an acceptable accuracy (85.3%) when identifying buildings that had blue tarps present on their roofs. Common confusion with the blue tarp class included high reflectance surfaces (e.g., metal roofs) and pools. Further refinement of the model could improve the accuracy and the introduction of other machine learning and/or object oriented classification methods may allow for detection of damage beyond tarps (Miura et al. 2020).

Filtering polygons by removing those with an area under 200 square feet did successfully reduce the noise in the model output. Researchers compared the accuracy between the spatially filtered and unfiltered results using the same methodology. The spatial filter improved the overall accuracy by approximately 7% and had a large impact on improving the model specificity. Previous studies have used object size to reduce classification errors (Jiménez-Jiménez et al. 2020; Palmer et al. 2018). Similar to the current study, these studies reduced noise by eliminating areas too small to be the roof in a primary structure. The improvement indicates the filter was appropriate for reducing the number of false positives in the final model results.

In addition, our results did indicate that some damaged homes were located in counties not eligible for individual assistance from FEMA. While the proportion was small (0.3–2.9%), the results still suggest there are households with damaged structures located in areas ineligible for assistance. To further investigate the pre-hurricane blue tarp prevalence, we counted the number of blue tarps present in Bay (FEMA A) and Jefferson (FEMA B) counties using data from the National Agriculture Imagery Program (NAIP). The NAIP images were collected in 2017 for Florida, approximately one year before the hurricane. The results showed low percentages of tarp in Bay (0.07%) and Jefferson (0.02%) counties, which are significantly less than after the hurricane (Bay 20.5% and Jefferson 0.9%). These findings show that while there may be blue tarps present before the storm, it is likely most were installed post-storm. In addition, the results provide evidence that there are areas in FEMA B that sustained damage. Media reports suggest that FEMA may be more likely to deny aid requests in rural areas struck by local disasters. Yet, aid is approved in statewide disasters, even if the damage they cause is less severe (Harris and Eaton 2022). This study provides an objective way for FEMA to target IA to sub-county areas that suffered infrastructure damage. If there is no state aid available, homeowners have few ways to seek financial aid.

Furthermore, there is evidence that communities that were economically disadvantaged and/or hosted a higher proportion of people of color were more likely to have aid requests rejected, which further compounds existing inequalities (Grube et al. 2018; Domingue and Emrich 2019). Previous studies have found evidence of pre-storm inequalities in that racial/ethnic minorities and populations with lower socioeconomic status are more exposed to inland flooding (Collins et al. 2019). Furthermore, these same populations are often living in areas with more poorly maintained infrastructure, older housing that may be built to less stringent building code requirements, and face lengthier recovery times (Peacock 2014; Wyczalkowski et al. 2019). Future work monitoring damage at the household level may provide further insight into inequities in storm damage and recovery and the impacts living in more severely damaged areas may have on health (Pan et al. 2021).

Blue tarps are distributed to qualified homeowners to provide temporary covering, protecting structures from further damage while repairs are being made. While the tarps are only meant to last a few months, studies have found the coverings are present from two to three months to over two years after a disaster (Rathfon et al. 2012; Rowan and Kwiatkowski 2020). There are various reasons why individuals are unable to repair their homes within the timeframe blue tarps are designed to last and for the most part, the most vulnerable face the greatest challenges. For example, although federal aid is available, it may be denied for failure to prove home ownership or issues during the home inspection process. In addition, even if approved for funds, in some cases the monetary amount can be insufficient to repair the home (García 2022). Failure to repair or have the home covered in the interim can result in further damage to the home (Allen 2007).

The study results have implications for health and indoor environments and methods can be applied for investigating these environments after disasters. Unsafe housing has a number of negative implications, including towards human health such as exposure to mold, mental health impacts, and chemical exposure. Studies conducted after Hurricane Katrina have examined the effects of dampness in the home on human health. Researchers found respiratory symptoms were associated with exposure to water damaged homes during clean-up activities in New Orleans following Katrina (Chulada et al. 2012; Cummings et al. 2008; Grimsley et al. 2012; Mitchell et al. 2012). In addition, flooding from hurricanes and debris removal can expose residents to harmful chemicals. After Hurricane Katrina residents were concerned about chemical exposures in air, soil, and water (Picou 2009). Lastly, studies have consistently found worsened mental health associated with disasters (Goldmann and Galea 2014). Although these health effects are linked to direct exposures such as injury, the long-term impacts of the disaster, such as loss of property, can lead to high levels of stress. Prolonged stress is associated with some of the most common mental health conditions after disasters, including posttraumatic stress disorder and major depressive disorder (Foa et al. 2006; Nillni et al. 2013; North and Baron 2021). Therefore, alleviating the stress associated with damaged housing has implications for improving mental health.

Limitations

This study has some limitations. First, we did not account for blue tarps in the entire study area that may have been present before Hurricane Michael struck Florida. However, we did provide estimates in two counties using NAIP imagery taken before the hurricane. Second, the images used to classify blue tarps were taken between two and four months after the hurricane made landfall. Some homes may have been repaired within this timeframe. In addition, for blue tarps to be installed by the Army Corps of Engineers, the roof must have less than 50% structural damage (FEMA 2017). Therefore, any residences too heavily damaged to install blue tarps are not considered through this classification. Third, blue tarps not located on roofs (e.g., blue tarps used to cover sheds or vehicles) would also be identified by the model. These tarps were largely excluded by the spatial filter that removed any areas under 200 square feet. Additionally, by limiting the analysis to occupied residential parcels, we excluded parcel land use codes without homes such as common areas and vacant residential properties. While we excluded these properties in an effort to limit inclusion of non-residences, it is difficult to distinguish multi-building damage within a parcel; for instance, if the non-primary structure on the property was damaged. Object-based classifiers may improve the accuracy of detecting blue tarp and other forms of roof damage.

Lastly, the study results did not include the severity of the damage. With building footprint data, it may be possible to provide the proportion of the damage for each structure. However, by relying on blue tarp presence alone, it is not possible to assess the severity of damage. Instead, blue tarp presence provides a practical assessment of damage to help frontline responders direct resources efficiently and effectively. Introducing a CNN and additional data (e.g., building footprints) may improve the accuracy as well as provide additional information about the damage sustained (Berezina and Liu 2022). CNN could also make use of spatial information and account for information in nearby pixels (Hasan et al. 2019). In addition, using image segmentation methods may help to improve the accuracy of the classification particularly when using such high-resolution data. Deep learning methods that can make use of the rich spatial information in the data may improve the accuracy of the model and eliminate some of the class confusion (e.g., confusion between blue tarps and pools) (Qi et al. 2020).

Conclusion

This study presents evidence that high resolution imagery provides a feasible method to detect post-disaster damage and recovery. Remotely sensed imagery has the advantage of being able to cover a wide area relatively quickly at relatively low cost compared to ground surveys. However, computing abilities and storage may limit the applicability of high-resolution data. Spatially filtering the model output improved the overall model accuracy and specificity. While this study focused on identifying blue tarps in counties affected by Hurricane Michael, future studies could expand this methodology to include more frequent data collection and deep learning methods to detect roof damage more generally and track recovery over time. Detection of damage in areas outside of those eligible for individual assistance also has the potential to provide further evidence of inequitable aid distribution. Safe housing is a basic need for disaster survivors and denying or providing inadequate aid has long-term implications for community resilience and recovery (He et al. 2021; Lichtveld et al. 2020; Shultz and Galea 2017). Remote sensing provides a means to assess damage quickly and monitor recovery in the months after a disaster to understand where disaster effects persist.