Multispectral satellite imagery provides a spaceborne perspective on tornado damage identification; however, few studies have explored how tornadoes alter the spectral signature of different land-cover types. In part 1 of this study, Landsat surface reflectance is used to explore how 17 tornadoes modify the spectral signature, NDVI, and “Tassled Cap” parameters inside forest (N = 16), grassland (N = 10), and urban (N = 17) land cover. Land cover influences the magnitude of change observed, particularly in spring/summer imagery, with most tornado-damaged surfaces exhibiting a higher median reflectance in the visible and shortwave infrared, and a lower median reflectance in the near-infrared spectral ranges. These changes result in a higher median Tasseled Cap brightness, lower Tasseled Cap greenness and wetness, and lower NDVI relative to unaffected areas. Other factors affecting the magnitude of change in reflectance include season, vegetation condition, land-cover heterogeneity, and tornado strength. While vegetation indices like NDVI provide a quick way to identify damage, they have limited utility when monitoring recovery because of the cyclical seasonal vegetation cycle. Since tornado damage provides an analogous spectral signal to that of forest clearing, NDVI is compared with a forest disturbance index (DI) across a 5-yr Landsat climatology surrounding the 27 April 2011 tornado outbreak in part 2 of this study. Preoutbreak DI values remain relatively stable across seasons. In the five tornado-damaged areas evaluated, DI values peak within 6 months followed by a decline coincident with ongoing recovery. DI-like metrics provide a seasonally independent mechanism to fill the gap in identifying damage and monitoring recovery.
a. Tornado damage surveys
In the wake of a tornado, ground and aerial surveys conducted by National Weather Service (NWS) personnel and partners provide a method to document the extent and intensity of the damage for use within studies regarding infrastructure resiliency (e.g., Marshall 2002; Federal Emergency Management Agency 2012), relationships between radar remote sensing and tornado strength (e.g., Toth et al. 2013; Kingfield and LaDue 2015; Smith et al. 2015), associations with tornado structure and dynamics (e.g., Davies-Jones et al. 1978; Fujita 1989; Karstens et al. 2013), and assessments of risk and vulnerability (e.g., Brooks et al. 2003; Ashley 2007). The enhanced Fujita (EF; Wind Science and Engineering Center 2006) scale, introduced in 2007 to mitigate deficiencies with the original Fujita (F; Fujita 1971) scale (e.g., Minor et al. 1977; Doswell and Burgess 1988), provides guidance on estimated wind speed and destruction relationships for 28 damage categories and is the current standard for tornado surveys in the United States. While there is a demand to establish evaluation consistency for each survey (Edwards et al. 2013; Burgess et al. 2014), many surveys are incomplete or rushed because of the time-sensitive and intensive nature of surveying coupled with workforce, training, and funding limitations (Doswell and Burgess 1988; Doswell et al. 2009). This leads surveyors to determine where the maximum damage occurred (Speheger et al. 2002), focusing on anthropogenic structures that make up 26 of the 28 EF-scale damage categories and often omitting nonurban or rural areas. While not a priority for storm surveys, damaged vegetation from natural hazards can influence radiative energy budgets and temperatures (Parker et al. 2005; Segele et al. 2005; McPherson 2007) and remain visible long after the initial event (Dyer 1988; Klimowski et al. 1998).
b. Multispectral satellite imagery
The use of spaceborne multispectral imagery has been encouraged (e.g., Bentley et al. 2002; Yuan et al. 2002; Jedlovec et al. 2006; Molthan et al. 2014) to supplement the information provided by damage surveys. Multispectral remote sensing provides a synoptic look at Earth’s surface by measuring radiation emitted or reflected from Earth and captured by the satellite’s multiple spectral bands. Each spectral band corresponds to a range of frequencies located along the electromagnetic spectrum and can span beyond the visible wavelength region, where human vision is constrained, into the near-infrared (NIR) and shortwave infrared (SWIR) wavelength regions.
The amount and spectral distribution of energy reflected by Earth’s surface is dependent on the surface characteristics with vegetation, soils, and anthropogenic materials (Fig. 1) providing unique spectral signatures received at the sensor. In the case of healthy vegetation (Fig. 1a), plants will absorb more energy in the visible (0.4–0.7 μm) region while reflecting more energy in the NIR (0.75–1.35 μm) to aid in photosynthesis. As a plant senesces, the opposite behavior will occur in the spectral reflectance curve. Soils (Fig. 1b) can have a variable spectral signature that is dependent on their structure, particle size, and organic/mineral composition. Many soils are highly reflective in the SWIR region (1.5–2.5 μm); however, wet soils tend to be less reflective in this region compared to dry soils (Gao 1996). Anthropogenic materials (Fig. 1c) also provide a distinct spectral reflectance curve that is highly dependent on its construction composition.
Damaging winds and debris associated with tornadoes will alter the physical and spectral signature of Earth’s surface. The damage magnitude is dependent on several factors including the tornado strength, size, and makeup of the underlying land cover within the tornado swath (Jedlovec et al. 2006). For example, the scattering of anthropogenic materials or scouring of the ground over areas of vegetation could disrupt the photosynthesis process, altering the difference in reflectance between the visible and NIR regions and enhancing reflectance in the SWIR region.
c. Spaceborne ratio-based analyses of thunderstorm damage
Initial observational studies of thunderstorm damage were often built around cases that were convenient because of relatively clear imagery received before and after the hazardous event. Many studies rely on change detection (Singh 1989) and use spectral vegetation indices, such as the widely adopted normalized difference vegetation index (NDVI), to identify the damage extent (Bentley et al. 2002; Yuan et al. 2002; Henebry and Ratcliffe 2003; Parker et al. 2005; Segele et al. 2005; Jedlovec et al. 2006; Wilkinson and Crosby 2010; Gallo et al. 2012; Wagner et al. 2012; Molthan et al. 2014). NDVI [Eq. (1)] compares the reflectance of the visible red and NIR spectral ranges (Rouse et al. 1973; Tucker 1979):
NDVI exploits the pattern that photosynthetically active vegetation reflects a higher fraction of radiation from the NIR and absorbs more radiation from the red wavelength regions, resulting in a higher NDVI value. During ongoing plant senescence, the reflectance values of these two bands will be closer together, resulting in a lower NDVI. Since NDVI is an optical measurement of vegetation health, this ratio-based index can be influenced by atmospheric and viewing geometry effects (Bannari et al. 1995) and is sensitive to the background canopy (Huete and Jackson 1987; Huete 1988). Newer vegetation indices have been developed to manage these limitations (Bannari et al. 1995), but the spectral ranges required to calculate them may not be available across different sensor platforms. NDVI is subsequently viewed as a “continuity index” because of its availability through several generations of satellites and relevance to a collection of environmental monitoring studies (Huete et al. 2002). While rooted in studies of plant phenology, NDVI has been uniformly applied to both vegetated and urban/agricultural cover types in prior analyses of thunderstorm damage (e.g., Yuan et al. 2002; Jedlovec et al. 2006; Wagner et al. 2012).
For all orbiting multispectral sensors, there is an inverse relationship between spatial and temporal resolution. To assist in postevent response efforts, a few studies have sacrificed spatial resolution for update frequency by using sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS; e.g., Jedlovec et al. 2006; Wilkinson and Crosby 2010). Jedlovec et al. (2006) compared 250-m-spatial-resolution MODIS imager data onboard the NASA Aqua and Terra satellites to higher-spatial-resolution imagery and found that MODIS-derived NDVI data may be sufficient at detecting tornado tracks rated F2 or higher, subsequently providing the ability for a timely assessment of damage extent; however, damage signatures appeared more clearly in areas of dense or homogenous vegetation (higher prestorm NDVI values), while damage was harder to distinguish in sparse or heterogeneous vegetation (lower prestorm NDVI).
The MODIS red and NIR spectral bands are the only bands with 250-m spatial resolution, with the other 34 bands at 500 m or 1 km. This limits the creation of other vegetation indices beyond NDVI at an equivalent resolution. Evaluating all tornadoes within the Storm Data database from the National Centers for Environmental Information (NCEI) since the inception of the EF scale (2007–15), we found that 17.3% of the tornado records had an estimated maximum width exceeding 250 m. This is based upon the assumption that all database records are accurate, which may not be the case (e.g., Witt et al. 1998). Regardless, a large majority of these tornadoes may not be observed using MODIS or sensors with similar spatial resolutions.
To address this limitation, other studies have used higher-spatial-resolution imagery (≤30 m) to discriminate between damage and nondamage areas. Estimating the amount of agricultural loss due to thunderstorm damage on 12 August 1999, Bentley et al. (2002) compared NDVI calculated using Landsat-7 Enhanced Thematic Mapper Plus (ETM+) imagery and observed swaths of lower NDVI values where large hail occurred. Evaluating multispectral returns across a heterogeneous terrain, Yuan et al. (2002) used the Linear Imaging Self-Scanning III 23.5-m-spatial-resolution imagery to generate NDVI postevent and change products over a damage swath from the 3 May 1999 Oklahoma City, Oklahoma, F5 tornado. They found spatial collocation between lower NDVI and F2+ tornado damage rated with some signal corresponding to F1 damage in rural areas. In another analysis of this tornado, Wagner et al. (2012) used Landsat-5/7 imagery to calculate NDVI alongside other indices and assess damage recovery over a 3-yr period. Recovery rates were mainly influenced by the amount and severity of the initial damage, with the hardest hit regions never completely recovering.
d. Analyses of disturbance
In other disciplines, disturbance identification is one of several techniques used to monitor and model ecosystem attributes (e.g., carbon emissions) within the larger Earth system (Cohen and Goward 2004). Forests damaged by natural hazards, human development, insects, and disease emit more carbon into the atmosphere, while new and recovering forests pull carbon from the atmosphere (Odum 1969). Kauth and Thomas (1976) showed that all four spectral channels on the Landsat-1 Multispectral Scanner (MSS) contained relevant information to monitor vegetation health. By weighting the sums of the MSS bands, they derived “Tasseled Cap” indices of brightness and greenness. With the launch of the Landsat-4 Thematic Mapper (TM), the Tasseled Cap indices were redefined to use six TM spectral bands and expanded to add wetness derived from digital numbers (Crist and Cicone 1984) and surface reflectance (SR; Crist 1985). Tasseled Cap indices, particularly wetness, are a valuable predictor of forest structural attributes (Cohen et al. 1995) and respond to the amount of green vegetation regardless of the background soil reflectance. While the origin of Tasseled Cap is rooted in agricultural assessment, its utility has been explored for other land-cover regimes, including urban regions (Crist and Cicone 1984; Deng and Wu 2012). From a data storage standpoint, a reduction in the number of variables stored from six Landsat spectral bands to three Tasseled Cap parameters have further promoted its usage for monitoring ecosystem disturbance across longer time scales (e.g., Cohen et al. 2002).
Capitalizing on the ability of Tasseled Cap to monitor ecosystem changes, particularly in forests, Healey et al. (2005) introduced a disturbance index (DI) technique that incorporates and reduces the three Tasseled Cap indices down to a single variable that highlights forest disturbance. Since its inception, DI has been used to describe decadal change in North American forests (Masek et al. 2008), fused with multiresolution satellite data (Hilker et al. 2009; Tran et al. 2016), and modified to evaluate grazing in New Zealand grasslands (de Beurs et al. 2016). Related back to thunderstorm hazards, Baumann et al. (2014) used DI alongside a set of spectral and Tasseled Cap thresholds to distinguish, with greater than 75% accuracy, forest pixels that were disturbed by windfall. Their success with wind damage promotes the applicability of DI in identifying areas of forest damage by tornadoes, which tends to be a more localized swath of defoliation and canopy damage; however, there has been little formal investigation on the application of DI in the identification of tornado damage to date.
In this study, we expand upon the existing knowledge base of tornado damage identification in two ways. In the first part of this study, we use Landsat-5 TM and Landsat-7 ETM+ imagery to explore how tornadoes change the spectral reflectance curves of forests, grasslands, and urban environments across different geographic regions in the United States.
Tree damage by tornadoes, particularly in forested regions, accounts for 2 of the 28 damage categories in the EF scale yet is a low priority for damage survey teams. For the second part of this study, we compare the DI technique of Healey et al. (2005) to NDVI in the immediate and longitudinal identification of tornado damage in forests through a 5-yr climatology of Landsat imagery surrounding the 27 April 2011 tornado super outbreak, where over 199 tornadoes occurred across the southeastern United States in one day (Knupp et al. 2014).
2. Data and study region
a. Landsat data
We acquired Landsat-5 TM and Landsat-7 ETM+ SR data from the U.S. Geological Survey (USGS) EarthExplorer platform (https://earthexplorer.usgs.gov).1 The USGS uses the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS; Masek et al. 2006) to atmospherically correct, quality control, and create the SR products. This correction mitigates atmospheric effects (i.e., aerosols and other radiative scatterers) and provides a more accurate estimate of solar radiation reflected by Earth’s surface compared to uncorrected, top-of-atmosphere measurements. Furthermore, a new cloud masking algorithm (CFMask) was added to LEDAPS to identify cloudy, cloud-adjacent, and cloud-shadowed pixels. This algorithm performs better at identifying these features than the legacy masking algorithm (Zhu and Woodcock 2012) and was used in this study to remove all pixels not classified as “clear.”
b. Land-cover data
The National Land Cover Database (NLCD; Homer et al. 2015) provides a 30-m-spatial-resolution grid of 20 land-cover classes derived from the classification system of Anderson et al. (1976). Since 2001, this database has been updated every five years with newer versions released for 2006 and 2011. In this study, we applied the most recent NLCD database prior to each tornado date to classify the land cover. For forest identification, we selected the deciduous (class 41), evergreen (class 42), and mixed (class 43) classification categories. For grassland identification, we only used the grassland/herbaceous (class 71) category as the other three herbaceous categories are native to Alaska. For urban identification, we used all four developed classifications (classes 21–24) ranging from open space to high intensity. While NLCD provides classification for pasture/hay (class 81) and cultivated crops (class 82), we did not evaluate these agricultural land-cover types because of the interseason variability of harvest cycles.
c. Case selection and study domain
NCEI maintains a tornado database dating back to 1950 with most tornado records containing estimated start and end times and location of occurrence. In part 1 of this study, we selected tornadoes with Landsat TM–ETM+ imagery within 30 days of occurrence over a diverse range of geographic regions in the United States. Imagery with visible ground scouring and previously studied tornadoes were prioritized as they corroborate with the storm event and literature record. In total, we selected 17 tornadoes (Table 1) occurring within 12 Landsat image tiles (black squares; Fig. 2).
In part 2 of this study, we focus on using DI to assess damage recovery in forested regions. To standardize the image times across the 5-yr study period, we selected five tornadoes (purple polygons; Fig. 2) with visible ground scouring from Landsat path 21/row 37 imagery covering portions of west-central Alabama and eastern Mississippi. For this path/row, we downloaded all Landsat-5/7 SR imagery from 1 January 2009 to 31 December 2013. Since DI was originally created for forest damage, we used the 2011 NLCD dataset coupled with the CFMask product to determine the percentage of clear-sky forest pixels in each image. Images with ≥10% cloud cover over forests were discarded. The Landsat-7 scan line corrector (SLC) failed after May 2003 and results in a 22% reduction in data coverage (Markham et al. 2004). Forest pixels within these missing data regions were treated like cloud-masked pixels. In total, 19 Landsat-5 and 30 Landsat-7 images were available for data analysis. Since the Landsat-5 mission ended in November 2011, all remaining images over the 5-yr period were from Landsat-7.
a. Damage and background pixel identification
We used the tornado record from NCEI alongside methods used in prior studies, such as the calculation of vegetation indices (Jedlovec et al. 2006; Wagner et al. 2012) and principal components analysis (Yuan et al. 2002; Molthan et al. 2014), as guidance to manually contour a damage polygon, as shown in Fig. 3 for the Jasper County, Missouri, tornado (tornado J). With many modern tornado records only providing start and end locations along with a maximum length and width of unknown accuracy, manual contouring standardizes the damage identification process and allows for a nonuniform 2D region to be identified.
To assemble a dataset of pixels unaffected by the tornado (referred to as background), we calculated a spatial buffer 10–60 km away from the damage polygon. This buffer size ensures a sufficient number of pixels exist for resampling even if the tornado track was near the edge of an image. To ensure the smaller damage and larger background regions have the same sample size and land-cover distribution, we randomly sampled the background region without replacement. Imagery with multiple tornadoes or other thunderstorm hazard damage (e.g., hail streaks) had these other regions masked out using a 10-km buffer during the background resampling procedure.
b. Calculation of NDVI and Tasseled Cap indices
1) Individual bands and NDVI
The 30-m bands of Landsat TM–ETM+ SR (bands 1–5 and 7) served as 6 of the 10 inputs evaluated in part 1 of this study. These six bands, each consisting of a small spectral range, together provide the spectral reflectance of targets from the visible to SWIR wavelength regions (Table 2). Because of the commonalities in spectral range between the sensors, we will hereafter refer to individual Landsat bands by their center wavelength or band category independent of the sensor. NDVI was calculated by using the SR values from the red (0.66 μm) and NIR (0.83 μm) bands [Eq. (1)].
2) Tasseled Cap indices
Crist (1985) derived the transformation coefficients (Table 3) for the visible to SWIR reflectance bands using the Landsat-4 TM to calculate Tasseled Cap brightness, greenness, and wetness, which are the foundation for calculating DI (Healey et al. 2005). These coefficients were developed with ground measurements to mitigate errors from atmospheric effects. Given the similarities in the spectral ranges between Landsat-5 and Landsat-7 (Table 2) coupled with the use of LEDAPS to atmospherically correct all imagery, we use the coefficients from Crist (1985) to calculate the Tasseled Cap indices for both satellite platforms. The interpretation of the Tasseled Cap indices is dependent on the targets being analyzed (Crist and Cicone 1984). Brightness is the weighted sum across six input bands and represents the overall reflectance. Greenness is the contrast between the NIR (0.83 μm) and the visible bands (0.49–0.66 μm) for vegetation identification and monitoring. Wetness contrasts the SWIR (1.67–2.24 μm) with the visible/NIR wavelength regions to quantify soil moisture content and vegetation density. While the Tasseled Cap has origins in monitoring vegetation response, these indices have also been used to discriminate vegetation, high-albedo regions, and low-albedo regions in urban areas (Deng and Wu 2012).
3) DI calculation in forests
For part 2 of this study, we use the Tasseled Cap indices above and follow the methods of Healey et al. (2005) to generate a DI image. The DI calculation uses a linear combination of normalized Tasseled Cap indices (Brightnessn, Greennessn, and Wetnessn) to determine how far, in terms of standard deviation, each individual forest pixel is from the global mean of forests of the same species (or land-cover type). In our study, we segment each image into one of the three NLCD forest classes (i.e., deciduous, evergreen, and mixed) and calculate an imagewide mean (μ) and standard deviation (σ) for each Tasseled Cap parameter and forest class. At each pixel, we then standardize each local Tasseled Cap parameter by the imagewide μ and σ associated with its forest class and parameter [e.g., as shown for Brightnessn in Eq. (2)]. Finally, these three standardized values are combined to calculate DI [Eq. (3)]. This process is repeated for all forest pixels in an image:
The DI equation is based upon the assumption that clear-cut forests will exhibit an increase in brightness and decrease in greenness and wetness compared to surrounding undisturbed forests. The DI value is interpreted as the spectral distance from that pixel to a mature pixel of the same forest class. Interpreting the change in DI over time shows that large positive shifts correspond to disturbance, while large negative shifts correspond to regrowth (Healey et al. 2005; Masek et al. 2008; Tran et al. 2016).
Evaluating disturbance in three forests, Healey et al. (2005) observed mean DI values ≥ 2 in areas affected by disturbance events. De Beurs et al. (2016) notes that setting a DI ≥ 2 as a threshold for disturbance can misclassify disturbed pixels 25% of the time. Setting higher DI threshold will lower the misclassification rate, but could potentially eliminate identification of lower-magnitude disturbance events. Baumann et al. (2014) found that windfall regions were best identified at a DI threshold between 2.5 and 3. Incorporating the concerns of de Beurs et al. (2016), we have chosen a more conservative DI threshold of 3 in part 2 of this study.
In our exploration of DI, we are only interested in disturbance caused by tornadoes. To mitigate preexisting forms of disturbance from being misidentified as tornado damage, we identified all predisturbed pixels (DI ≥ 3) in the 2 April 2011 Landsat-7 image prior to the tornado outbreak. All forest pixels failing LEDAPS quality control in the 2 April 2011 image were rechecked in older Landsat images. After identification, all predisturbed pixels were removed from the calculations of DI and NDVI for the entire 5-yr period.
4. Results—Identifying tornado damage by land cover
Comparing the sample sizes of urban, forest, and grassland pixels affected by tornadoes (Table 4), we found that forests were the most affected land-cover type in 13 of the 17 tornadoes followed by urban areas (3) and grasslands (1). We removed tornadoes from an individual land-cover analysis with fewer than 100 pixels as the poor sample sizes skewed the distributions of SR, Tasseled Cap indices, and NDVI. After applying this threshold, we removed one forest case (tornado F) and seven grassland cases (tornadoes D, F, G, H, J, M, and Q in Table 4).
a. Urban land cover
All tornadoes had at least 100 damaged urban pixels ranging from 123 (tornado N) to 5530 (tornado J) corresponding to the EF5 tornado that devastated Joplin, Missouri, on 22 May 2011. Relative to the background region, all tornado-damaged urban areas had a higher reflectance in the visible (0.49–0.66 μm) and SWIR (1.67–2.24 μm), and most areas exhibited a lower reflectance in the NIR region (0.83 μm) (Fig. 4a). In the blue (0.49 μm) and green (0.56 μm) regions, the SR was around 15% higher, while red (0.66 μm) and SWIR (1.67–2.24 μm) values were around 25% higher in the damaged region. Changes in NIR (0.83 μm) SR were much smaller, with many tornado-damaged areas having around a 5% lower reflectance compared to the background region. The general increase in SR measured across most Landsat bands corresponded to a 25% higher median Tasseled Cap brightness in the damage polygon. With NIR reflectance holding the largest and only positive weighting coefficient in the greenness calculation (Table 3) coupled with many pixels registering a lower NIR reflectance, median greenness values were around 25%–60% lower in the damaged areas. The negative weights in the wetness calculation paired with higher SWIR SR inside the damage polygon results in median wetness being 25%–75% below the median wetness values in the background region. The altered spectral signature in tornado-damaged urban pixels also yields a decline in median NDVI because of higher reflectance at 0.66 μm and lower reflectance at 0.83 μm. For example, the tornado-damaged areas associated with tornado P were around 64% more reflective at 0.66 μm but around 1% less reflective at 0.83 μm. This contraction in the red edge between the 0.66- and 0.83-μm bands yields a median NDVI of 0.26 in the damage region compared to 0.48 in the background region. This corroborates well with other observations of decreases in NDVI due to tornadoes (e.g., Yuan et al. 2002; Jedlovec et al. 2006; Molthan et al. 2014).
The spectral reflectance of urban environments is dependent on the infrastructure density and is a contributing factor in evaluating the magnitude of tornado damage. In the case of the Jasper tornado (tornado J), pixels classified as open-space urban (Fig. 5a) saw a larger decline in median NDVI (−0.18) compared to high-density urban pixels (−0.06, Fig. 5b). Open-space urban areas are more likely to consist of a mixture of constructed materials and vegetation, while high-density regions are predominantly composed of impervious surfaces (Homer et al. 2015). As expected, Landsat TM–ETM+ measures a lower reflectance in the visible wavelengths and higher reflectance at 0.83 μm in the open-space region compared to the high-intensity regions likely corresponding to photosynthetic activity in these vegetated areas. After the tornado occurs, the likely addition of debris covering/shadowing the open-space vegetation coupled with an inconsistent defoliation of the vegetation area and displacement of dirt/soil increases reflectance in the visible and SWIR and decreases reflectance in the NIR spectral ranges. In high-density urban areas, the increased amount of anthropogenic material and impervious surfaces (Fig. 1c) results in a higher initial reflectance in the visible and SWIR wavelength regions. After the tornado, only the 0.66- and 2.27-μm regions showed increased reflectance between 3% and 5% above the background field, while up to an 8% decline in median reflectance was measured in the other Landsat spectral ranges. In the Tasseled Cap space, background high-density urban areas are brighter and have lower greenness and wetness values than open-space urban areas. Tornado-damaged open-space urban areas have a higher median brightness (0.07) and lower median greenness (−0.08) and wetness (−0.11) values, whereas high-density pixels do not deviate as far from the background in measures of central tendency. However, tornado-damaged pixels in both urban land-cover types have a smaller interquartile range (IQR) across the spectral reflectance curve and this corresponds to a lower point spread in the Tasseled Cap space.
One other observation of note is related to the potential seasonal effects on the spectral signature of cities. The Scott–Newton, Mississippi (tornado N), and Marshall, Mississippi (tornado Q), events occurred in the fall and winter months. In both events, NIR reflectance in tornado-affected urban land cover was 15% and 13% higher than the background region; however, red reflectance also was 38% and 24% higher, respectively. While both cases still resulted in a decline in NDVI, the density distribution of urban environments and background signal coupled with season can hinder tornado damage identification using automated techniques trained based on trends observed only in the spring and summer.
b. Forest land cover
Forest land cover showed the most consistent change across all parameters evaluated for the 16 valid tornado events. Comparable to urban area trends, tornado-damaged forests have a higher visible (0.49–0.66 μm) and SWIR (1.67–2.24 μm) reflectance and a lower NIR (0.83 μm) reflectance (Fig. 3b). Comparing damaged and background SR in tornadoes D (Fig. 6a), E (Fig. 6b), and G (Fig. 6c) shows these trends are geographically independent with documented occurrences in Wisconsin, Missouri, and Mississippi, respectively. In these three cases, the SR of tornado-damaged forests was around 25%–69% higher in the blue/green (0.49–0.56 μm), 70%–134% higher in the red (0.66 μm), 17%–34% lower in the NIR (0.83 μm), 40%–65% higher in the SWIR (1.67 μm), and 79%–144% higher in the SWIR (2.24 μm) spectral regions. Defoliation, broken limbs, and downed trees will reduce the amount of photosynthetic vegetation and subsequently result in higher reflectance values in the visible and lower reflectance in the NIR spectral regions; the opposite spectral behavior of a region with healthy vegetation. Reflectance in the SWIR wavelength region is inversely related to the amount of moisture in vegetation (Schroeder et al. 2011) and directly related to the soil background (Cohen and Goward 2004). The limited amount of healthy vegetation to retain moisture and higher probability of soil exposure due to canopy defoliation are two of the potential drivers for the increase in SR observed here.
In the Tasseled Cap space for these three tornadoes, we observe an increase in brightness and decrease in greenness and wetness in the region damaged by the tornado; this corroborates with other studies of forest damage (e.g., Cohen and Goward 2004; Healey et al. 2005; Baumann et al. 2014). However, these changes can be variable depending on the damage extent, as downed or stripped trees can leave shadows across the affected area that lower brightness and increase wetness when compared to regions where trees are completely removed (Baumann et al. 2014). A comparison of the IQRs between the two pixel areas reveals a larger spread in the SWIR wavelength region. This is likely due to contributions of increased visibility of the soil background coupled with vegetation damage compared to healthy forests in the same geographic area and also results in a wider range of wetness values measured in the Tasseled Cap space. The IQRs in the visible and NIR wavelengths between the damaged and background domains are volatile and ultimately dependent on both the species of vegetation affected, position in the natural vegetation growth–decay cycle, and magnitude of damage that can fluctuate across the tornado life cycle (Holland et al. 2006).
Alongside the disruption of the photosynthesis process and exposure of the soil background in forests, another factor that can influence SR returns in tornadoes is the displacement of debris from one land-cover region onto another. In both the Cleveland, Oklahoma (tornado P; Fig. 7a), and Jasper (tornado J; Fig. 7b) tornadoes, forest pixels are dispersed around expansive urban areas that comprise 94.6% and 76.9% of the damage polygon area, respectively. Median SR values of forests were between 110% and 242% higher in the visible, 88% and 186% higher in the SWIR, and around 6% lower in the NIR spectral ranges compared to the background. While the reflectance departures follow the same pattern as other forest analyses (e.g., Figs. 5a–c), the collocation with the urban environment allows for the scattering of anthropogenic debris and could be an additional factor driving reflectance upward in the visible and SWIR wavelengths.
c. Grassland land cover
Comparable to urban and forested regions in terms of the directional change in reflectance, most tornado-damaged grasslands registered median reflectance values around 15%–25% higher in the visible and SWIR and 5% lower in the NIR (Fig. 3c). The Canadian–Kingfisher–Logan, Oklahoma (tornado L), event produced damage in a rural portion of northwest Oklahoma and contained the highest number of damaged grassland pixels (N = 11 158). Five days after the tornado, median SR values of tornado-damaged grasslands are around 24%–44% higher in the visible (0.49–0.66 μm) and SWIR regions (1.67–2.64 μm) and slightly lower with a 2% decrease in the NIR region (0.83 μm) compared to the background region (Fig. 8a). The decline in NDVI tends to be more dependent on the red (0.66 μm) SR as this departs further from the background over the NIR reflectance. Comparable to other land-cover types, these spectral signatures from the background resulted in an increase in median brightness and decrease in greenness and wetness.
Similar to observations of damaged forests collocated within urban areas, grasslands around the EF5 tornado affecting Cleveland (tornado P; Fig. 8b) had greater differences in SR between the damaged and background regions. SR values were 37%–90% higher in the visible and SWIR and around 1% lower in the NIR spectral ranges. The increase in red reflectance translates to a lower median NDVI (0.38) compared to the background (0.62). These larger departures in median reflectance between the damaged and background regions also translates to the greatest differences in the Tasseled Cap indices with an increase in median brightness (0.11) and a decrease in greenness (−0.07) and wetness (−0.10). Tornado N was the only grassland tornado sampled in the fall region and showed a 3% increase in the visible blue (0.49 μm) and a decline in all remaining bands ranging from 2% to 17% compared to the background region (Fig. 8c). In this case, the Tasseled Cap indices show the opposite of most other tornado damage tracks with a decrease in median brightness (−0.02) and increase in median greenness (0.01) and wetness (0.04). While the median NIR reflectance is lower than the background region, the distribution of values is negatively skewed and results in a higher median NDVI measured in the damaged region. Similar to what was seen for forests, the IQR of each spectral region does not follow a common trend across different tornadoes. In the case of tornado L (Fig. 8a), the IQR of each Landsat band is smaller in the damage region, while the opposite is true in tornado N (Fig. 8c). The amount and extent of damage produced by the tornado, vegetation strength, soil type, season, and debris lofting from surrounding land cover are several factors that can change the spectral behavior of vegetated areas. Furthermore, unlike forests and urban land cover, spectral signatures of grasslands are more susceptible to external influences (e.g., drought, grazing, land management practices; Turner et al. 1992; de Beurs et al. 2016) and can hinder the identification of damage. Jedlovec et al. (2006) observed that manual and automated identification of damage signatures was easier when evaluating areas of dense vegetation (e.g., standing trees), while regions of open grassland provided a lower detection efficiency.
5. Tracking damage recovery over time: NDVI versus DI
In part 1 of this study, forests composed a majority of the contoured damaged pixels in 47% of the tornadoes evaluated, with most forests exhibiting an increase in median Tasseled Cap brightness and a decrease in greenness and wetness. This is analogous to a disturbance signal that would be visible using DI. The five tornadoes selected for part 2 occurred within the 27 April 2011 tornado outbreak with damage lengths ranging from 20.3 to 258.6 km, maximum widths from 307.3 to 2306.6 m, and forests constituting 41%–81% of the damaged area (Table 5; Fig. 2). In a Landsat-7 image acquired on 4 May 2011, seven days after the tornado outbreak, a comparison between the damage and background regions for all tornadoes shows a spectral signature similar to observations in forests across other locations. The extent of the damage as related to differences in reflectance between the damage and background regions is correlated to the intensity of the tornado. In the case of the Wateroak, Alabama, tornado (Fig. 9a), SR values were around 18%–50% higher in the visible, 11% lower in the NIR, and 17%–38% higher in the SWIR region. For the Tuscaloosa–Birmingham, Alabama, tornado (Fig. 9b), the widest tornado in our DI dataset, SR values were between 28% and 115% higher in the visible, 18% lower in the NIR, and between 39% and 78% higher in the SWIR spectral regions. Accordingly, these SR departures permeate into the calculations of NDVI and DI. For the Wateroak tornado, the damage region had a lower median NDVI of 0.81 and a higher median DI of 0.95 compared to the background region at 0.89 and −0.99, respectively. In the stronger Tuscaloosa–Birmingham tornado, the median NDVI (DI) inside the damage region was 0.72 (3.40) compared to the background region at 0.88 (−0.97).
A comparison of DI to NDVI across the entire 5-yr period for the smallest (Fig. 10a) and largest (Fig. 10b) tornadoes reveals two distinct patterns before and after the tornado outbreak. Before the outbreak, DI remains stable in both the damage and background regions, with the median DI remaining at or below 0 and the total percentage of disturbed pixels remaining below 10% for a majority of images. As expected, NDVI over the same period varies with the seasons, with median NDVI values peaking around 0.89 in the midsummer months and reaching a minimum around 0.53 in the winter months. Because of this known variability of NDVI across seasons, employing change-detection techniques can provide very different results depending on the amount of time between the images. For example, over the period between March and May 2010, median NDVI across the background region increases from 0.56 to 0.79. Over this same period, we also observe increases in Tasseled Cap brightness (0.08), greenness (0.14), and wetness (0.05) in conjunction with more widespread growth and coverage of foliage within these forests. However, since a global mean and standard deviation are calculated within each image, the resulting summary statistics of DI are seasonally independent. This process of standardizing based on the current global spectral reflectance is the biggest strength of DI and mitigates several sources of cross-temporal signal contamination (i.e., vegetation growth cycles, droughts) that would be observed in standard ratio-based indices such as NDVI.
A comparison of both indices after the tornado outbreak provides additional evidence supporting the usage of DI or the calculation of a DI-like index. For the Wateroak and Tuscaloosa–Birmingham tornadoes, the percentage of disturbed pixels escalates rapidly with 25% and 60% of the respective swaths classified as disturbed within 6 months following the tornado outbreak. Over time, both the percentage of disturbed pixels and median DI converge toward the background signal. Alternatively, there is a decline in median NDVI after the tornado outbreak that continues to fluctuate with the seasons. Similar to DI, NDVI will converge with the background signal, as observed with the Wateroak tornado in the latter part of 2013 (Fig. 11a).
To bypass the dependence on season with indices like NDVI, prior multiyear studies of tornado damage have solely used data from within the same season (e.g., Wagner et al. 2012). DI has utility in allowing for imagery to be pulled from any season and assembled into a time series, providing a way to track and monitor damage year-round. For four of the five events (excluding Wateroak), over 45% of the pixels were classified as disturbed within 6 months after the tornado outbreak and never returned to pretornado disturbance levels at the end of December 2013 (Fig. 10a) whereas NDVI is unable to provide as clear of a recovery trend. However, the quality of DI is heavily dependent on the sample size coupled with the accuracy of the land-cover classification of the input imagery. Cloud cover can obscure forest pixels and portions of the damage swath, affecting both the global mean and standard deviation of the Tasseled Cap indices and summary statistics related to the disturbed pixels. This can be observed when relaxing the cloud-free threshold from 90% (Fig. 11a) to 50% (Fig. 11b). For example, the percentage of disturbed pixels (median DI) from the Sawyerville–Eoline, Alabama, tornado varies from 49% (2.89) on 15 January 2012 to 16% (0.81) on 22 May 2012 and back up to 35% (1.79) on 2012 June 23. On 22 May 2012, roughly 31% of the damage swath was covered in clouds/cloud shadows, removing areas with higher DI values. Furthermore, our reliance on Landsat-7 ETM+ data with the SLC failure reveals/removes certain parts of the tornado swath and background forest pixels within each successive image and contributes to some of the interimage variability in DI. Applying a stricter cloud-free threshold and using an alternate satellite sensor without the SLC failure (e.g., Landsat-8) would mitigate these artifacts.
6. Discussion and operational relevance
Comparing the spectral reflectance across grassland, urban, and forest land cover reveals several commonalities in how the spectral behavior changes within tornado-damaged areas. Most affected areas had a higher SR in the visible and SWIR and a lower SR in the NIR Landsat spectral bands. In most tornado and land-cover combinations evaluated, the smaller difference between the NIR and visible red reflectance resulted in a decline in the median NDVI in tornado-impacted areas. These land-cover-independent observations coupled with the ability to calculate NDVI with relative ease support the use of vegetation indices such as NDVI to visualize the extent of tornado damage.
In this study, we were selective in which tornado swaths were included in our case dataset. In many instances, the manual contours were smaller than the spatial extent of tornado damage because of a lack of confidence in discriminating whether or not damage had occurred as a result of the spatial constraints of the tornado. In an analysis of satellite-estimated lengths versus surveyed lengths of tornadoes from the 27 April 2011 outbreak, Molthan et al. (2014) measured that satellite-based lengths underestimated survey lengths by around 20 km with Landsat-7 ETM+ data and 23 km with higher-resolution ASTER data. Regardless, satellite data can aid in identifying regions of damage potentially missed or unreachable by ground-based damage surveys.
It is worth repeating that attempts to detect tornado damage are dependent on a multitude of variables. First, the ability to identify a damage signature is subject to the spatial constraints of the phenomenon observed, be it a tornado, damaging wind, or hail event. The events in this study and many prior studies (Yuan et al. 2002; Jedlovec et al. 2006; Wilkinson and Crosby 2010; Molthan et al. 2014) showed the clearest damage signal had visible scouring observed in the SR imagery and was associated with very intense tornadic circulations. However, tornadoes go through varying levels of organization throughout their life cycle, producing inconsistent and oftentimes asymmetric damage in their wake. Holland et al. (2006) simulated how tornadoes damaged forests and found variable responses to tree fall that were dependent on both the radial and tangential components of the vortex and subsequent forward speed. These translational variables coupled with the geographical location, health, and composition of forests will affect the extent remotely sensed imagery can detect damage. A comparable argument could be made on the dependence of building construction quality and quantity in determining the extent of tornado damage in urban areas. A second factor affecting damage identification is the scale and contribution of different land-cover types in a geographic domain. Comparing urban densities, we found that high-density urban areas showed a much smaller change in NDVI compared to open-space urban areas. Additionally, vegetated areas collocated or downwind of an urban environment showed a different pretornado spectral reflectance curve and experienced amplified departures in reflectance compared to analyses of damage further away from urban environments. The displacement of debris, exposure of soil, and defoliation of trees fundamentally alters the spectral behavior of Earth’s surface; however, the exact contributions of each factor may not be readily known with spaceborne data in the absence of ancillary data. Finally, the seasonal effect in identifying tornado damage is an underexplored topic. Many prior studies and this current study have focused on the identification of spring and summer tornadoes (e.g., Yuan et al. 2002; Jedlovec et al. 2006; Molthan et al. 2014) where the frequency of tornado occurrence is higher (Brooks et al. 2003). In the case of forest damage associated with tornado Q in Marshall, the only winter tornado in our study, NIR reflectance was 15% higher and red reflectance was 46% higher in the tornado damage polygon, resulting in a decline in NDVI by 0.07. The increase in the NIR spectral range was more prevalent in fall–winter events and coupled with the factors above may limit detection outside of the spring and summer.
In the immediate aftermath of a tornado, vegetation indices like NDVI can provide a cursory look at where tornado damage has occurred; however, the vegetation cycle associated with the seasons severely limits the application of comparing NDVI on a pixel-by-pixel basis across multiple images. The application of DI to identify tornado damage in forests in section 5 highlights its resiliency to the shortfalls of NDVI by providing a seasonally independent technique to initially identify as well as track damage areas over time. With a relatively stable DI observed in the 28 months prior to the tornado outbreak, change-detection studies using DI or a DI-like analog may not need a pre-event image in the same season or year to compare to a postevent image. However, a longer comparison interval opens up the potential for other sources of disturbance (e.g., wildfire, harvest, land-cover change) to enter the image and be confused with tornado damage. In a similar vein to NDVI, examination of the shape and extent of the damage will assist in determining the source of the damage.
While DI is beneficial at standardizing images across time, it is similar to NDVI in terms of missing detections because of either weaker tornadoes or tornado occurrence in certain geographic areas. Masek et al. (2008) observed diminished skill in the DI parameter from samples taken within the Rockies and Intermountain West. An optimal initial operating condition for this iteration of DI is a dark, closed-canopy forest that minimizes the contributions from the low-level vegetation and soils from dominating the reflectance signal. This issue can be mitigated by incorporating higher-resolution maps of vegetation type and density, if available. Furthermore, DI was originally developed to identify stand-replacing clearance (i.e., complete removal of trees). In the case of tornado damage, the remnants of downed trees and their associated shadows could cover up a higher fraction of the soil background and subsequently provide a lower-magnitude increase (decrease) in the Tasseled Cap brightness (wetness) values compared to clear-cut forests (Baumann et al. 2014). As such, adjustments to the forest coefficients based on the indigenous vegetation of a local geographic region along with the derivation of DI-like coefficients could provide a greater operational applicability to this index.
In the first part of this study, we explored how the spectral behavior of tornado damage varies within 17 urban, 16 forest, and 10 grassland environments across the central and eastern United States using Landsat-5 TM and Landsat-7 ETM+ imagery. Overall, we found that most tornadoes exhibited higher reflectance in the visible and SWIR TM–ETM+ spectral ranges and a lower reflectance in the NIR spectral range, particularly for tornadoes evaluated during the spring and summer months. During these seasons, these spectral signatures correspond to many tornado-damaged regions having higher Tasseled Cap brightness values as a result of a general increase in reflectance across most Landsat bands, lower Tasseled Cap greenness values driven by the decline in NIR reflectance, and lower Tasseled Cap wetness values due to the larger increases in SWIR reflectance compared to visible reflectance. Additionally, median NDVI values were lower in tornado-damaged areas for the three land-cover types except for one grassland case. While these trends in spectral signatures, Tasseled Cap indices, and NDVI may provide some initial guidance at developing a damage identification algorithm, the extent and magnitude of tornado damage is dependent on factors related to the strength of the tornado, the type and density of land cover it passes over, and time of year.
While NDVI is beneficial at providing a cursory look at localized change caused by natural hazards, analyses of recovery using NDVI is limited to the acquisition of cloud-free, intraseason imagery. Even with these constraints, pixel-based NDVI values can vary within season because of the normal vegetation life cycle and other external influences (e.g., drought). Part 2 of this study explored the applicability of using DI from Healey et al. (2005) to identify and track tornado damage in forested areas within a 5-yr window surrounding the 27 April 2011 tornado outbreak. Before the tornado outbreak, the median value of DI remained relatively stable compared to NDVI, which increases in the spring and summer months and decreases in the fall and winter months. After the tornado outbreak, values of DI initially increased dramatically with four of the five tornadoes registering over 45% of the contoured track as disturbed (DI ≥ 3) within 6 months. The percentage of disturbed pixels declined over time as ongoing recovery occurred in the region. During recovery, NDVI values began to converge with NDVI values in the background region but continued to vary with the seasons. While the quality of the DI output is influenced by the accuracy of the land-cover classification and the amount of cloud-free imagery, this seasonal resiliency should promote the usage of DI and the redevelopment of new DI-like indices focused on tornado damage in different land-cover regions in future studies of natural hazards.
The authors thank Braden Owsley, Brandon Smith, and Sean Waugh for their assistance in a pilot study on this topic. Discussions with Kevin Manross and Christopher Karstens were helpful in refining the methods used in this paper. Manuscript reviews by Andrew Molthan and one anonymous reviewer significantly improved the content and organization of this paper. Funding was provided by the NOAA/Office of Oceanic and Atmospheric Research under NOAA–University of Oklahoma Cooperative Agreement NA11OAR4320072, U.S. Department of Commerce, and through the NASA Interdisciplinary Science Program project NNX12AM89G as part of the Grant “Storms, Forms, and Complexity of the Urban Canopy: How Land Use, Settlement Patterns, and the Shapes of Cities Influence Severe Weather.” The views expressed in this paper are those of the authors and do not necessarily represent those of the NOAA, NSSL, or CIMMS.
Bulk downloading of Landsat products is also available at http://espa.cr.usgs.gov/.