Detecting Invasive Amur Honeysuckle in Urban Green Spaces of Cincinnati, Ohio Using Landsat-8 NDVI Difference Images
Bridget Taylor (corresponding author: University of Cincinnati, Department of Geography, 401 Braunstein Hall, Cincinnati, OH 45221-0131, [email protected]), Denis Conover (University of Cincinnati, Department of Biological Sciences, Cincinnati, OH), and Richard Beck (University of Cincinnati, Department of Geography, Cincinnati, OH).
Lonicera maackii (Amur honeysuckle) is a predominant invasive mid-level shrub introduced to North America from Asia in the late 1800s as an ornamental shrub. It prohibits native plants from growing over large areas of hardwood forests in the Midwestern and Eastern United States (Luken and Thieret 1996). Lonicera maackii was first documented as an escaped invasive in Hamilton County, Ohio around 1960 (Braun 1961) and has successfully invaded many green spaces in southwest Ohio (Hutchinson and Vankat 1998). When it invades a forest it alters the natural patterns of forest succession, reducing the growth rate of overstory trees by more than 50% (Hartman and McCarthy 2008). Lonicera maackii outcompetes many native plants with its growth pattern and extended growing season (McEwan et al. 2009, Wilfong et al. 2009). Early detection of invasive species is important to urban forest management because it curtails their spread (Moody and Mack 1988). Remote sensing methods can be a valuable tool for increasing knowledge about spatial patterns and predictors of invasion (Bradley 2014). Natural vegetation in cities is essential for maintaining biodiversity; this is known due to the extensive studies on adverse effects of natural habitat loss and fragmentation (e.g., Saunders et al. 1991, Gardner et al. 1993, Opdam et al. 1995, With et al. 1997). Urban green spaces support wildlife including rare and endangered species (Gibson 1998, Mortberg and Wallentinus 2000) and provide environmental services (e.g., air filtration, micro climate regulation, noise reduction, rainwater drainage, and recreation) (Bolund and Hunhammer 1999, Elmqvist et al. 2004).
Due to the extended leaf phenology, L. maackii can potentially be separated from native vegetation using remotely sensed data. The foliage of green plants strongly absorbs wavelength energy in the visible red band (Red) and somewhat reflects near infrared (NIR) (Rouse et al. 1973). A ratio of the red to near infrared reflectance, known as Normalized Differential Vegetation Index (NDVI), provides a valuable index of the greenness of a vegetation landscape. Wilfong et al. (2009) demonstrated mapping of L. maackii using Landsat-5 Thematic Mapper (TM) satellite imagery and Landsat-7 Enhanced Thematic Mapper Plus (ETM+) imagery by creating a difference image from a January NDVI image and a November NDVI image. Combining Landsat images with medium to fine spatial resolution and recorded GPS known locations was shown to be an effective method for early detection of L. maackii in a rural setting (Johnston et al. 2012) and in an urban setting (Shouse et al. 2013).
Remote sensing of L. maackii in urban forests is lacking, even though there is growing evidence to suggest L. maackii grows extensively in areas that are more urbanized (Reichard and White 2001, Harris et al. 2009, White et al. 2014). Shouse et al. (2013) compared L. maackii imagery detection methods in Cherokee Park (Louisville, KY) between more expensive and accurate high spatial resolution (HSR) imagery to freely available medium spatial resolution (MSR) imagery. Landsat-5 MSR imagery (30 m2 spatial resolution) was less accurate but still beneficial for detecting invasive versus native species (Shouse et al. 2013). L. maackii is more pervasive in urban forests compared with rural ones (Borgmann and Rodewald 2005). Urban landscapes are often highly fragmented, irregularly shaped, and with multiple land uses making remote sensing more challenging compared to rural areas (Gong et al. 1992, Barnsley et al. 1993, Giner and Rogan 2012).
Our research objective was to develop an inexpensive and efficient L. maackii mapping approach for urban forest ecological restoration land managers to use when trying to plan restoration projects in critical urban environments. In this study, we tested the feasibility of using a Landsat-8 (30 × 30 m spatial resolution) NDVI difference image for November and January with known locations of L. maackii presence and absence to identify a threshold for L. maackii detection. We mapped the spatial distribution of L. maackii in urban green spaces with extensive present and absent patches in Cincinnati, Ohio in Hamilton County (Figure 1). The NDVI difference image for November and January captures the change in vegetation greenness from November, where only L. maackii is green, to January, where almost no vegetation is green. We describe a threshold technique to identify change/present L. maackii with a higher NDVI difference value compared to no change/absent where L. maackii is absent. We recommend future research to validate this proposed threshold technique with a separate dataset.
Study area: five urban green spaces in Cincinnati, Ohio.
The study areas consist of five urban green spaces in Cincinnati, Hamilton County, Ohio: Bender Mountain Nature Preserve, Buttercup Valley Nature Preserve, Parkers Woods, Avon Woods Nature Preserve, and an undeveloped wooded section of Spring Grove Cemetery (Figure 1, Table 1). Urban green spaces defined in this study are predominantly biological preserves (Protected Areas) comprised of hardwood forest surrounded by dynamic and interacting urban infrastructure. Only the undeveloped wooded section of Spring Grove Cemetery is not a Protected Area since the land will be used as burial sites in the future. However, this section of the cemetery still presently fits the green space criteria in every other way.
Study area urban green space site locations and ownership.
Starting in 2004, Western Wildlife Corridor volunteers eradicated L. maackii from most of Bender Mountain Nature Preserve, except for the south hillside slope due to the concern that L. maackii removal might cause landslides onto the buildings at the bottom of the hill (Conover and Sisson 2016). Park volunteers eradicated much of the L. maackii in Buttercup Valley Nature Preserve. Parkers Woods, connected to Buttercup Valley, remains full of L. maackii. Spring Grove Cemetery and Arboretum is 283.28 ha with 114.52 ha of undeveloped green space in the northern section of the property. Avon Woods Nature Preserve has various invasive plants including L. maackii and has had scattered eradication efforts.
We downloaded two relatively cloud-free Landsat-8 images for the study area from USGS’s Earth Explorer for November 24, 2015 and January 24, 2015. These Landsat-8 images were atmospherically corrected and converted to the standard USGS surface reflectance product (USGS 2019). The November image was captured when L. maackii was green and native woody vegetation was dormant. In the January image, both L. maackii and native woody vegetation were dormant. We loaded these two images, supplementary map files, and our global positioning system (GPS) ground observation data into a GIS using the WGS84 ellipsoid and UTM 16 North projection.
We used the raster calculator tool in ArcGIS to calculate NDVI using the formula, NDVI = ([NIR—Red] / NIR + Red) (Rouse et al. 1973). For Landsat-8, band 5 is the NIR and band 4 is the Red. The output is an image with values varying between –1 (e.g. water) and 1 (full vegetation greenness). Wilfong et al. (2009) found that an NDVI difference image where the January image was subtracted from a November image provided better prediction of L. maackii than the use of a single November NDVI image. Therefore, we subtracted the January 2015 NDVI image from the November 2015 NDVI image to produce an NDVI difference image showing where vegetation greenness had changed the most between November and January. We clipped the NDVI difference image to the boundaries of each study site.
We used a hand-held GPS to record binary classifications of L. maackii absence (eradicated in patches > 900 m2) and L. maackii presence (> 900 m2) in the center of each patch type. We conducted random searches throughout the study sites relying on the knowledge of Dr. Denis Conover and his extensive previous research in southwestern Ohio (Conover and Geiger 1993, Conover and Geiger 1999, Conover et al. 2016, Conover and Sisson 2016, Conover et al. 2017). To avoid spatial autocorrelation, we took GPS records at intervals larger than 30m. We conducted field sampling from July 2015 through September 2015, and March 2016. A total of 46 presence and 26 absence L. maackii GPS points were collected. In our GIS, we also classified the GPS observation points with the NDVI difference pixel value that contained the point.
We ran a binary logistic regression, summary statistics, and a histogram of the November and January NDVI difference values compared to the GPS observation points to threshold the difference images into two categories of “change/present” and “no change/absent.” The logistic regression between NDVI difference values (predictor) and GPS observation (response variable) suggests there is a significant difference between L. maackii present and L. maackii absent; p-value was < 0.0001 and R² (Nagelkerke) value was 0.79. The histogram shows that higher NDVI difference values were observed to be L. maackii and the median NDVI difference value was 0.035 (Figure 2). We used these results to threshold the NDVI difference image at 0.035, where values greater than 0.035 were classified as L. maackii present and values less than 0.035 were classified as L. maackii absent.
Histogram of November minus January NDVI difference values (bin = 0.005) corresponding with L. maackii absent GPS points and L. maackii present GPS points.
The threshold we proposed correctly matched 59 of 72 ground-based observations (82%) but varied by site (Table 2, Figure 3). The threshold accuracy rate varied from 50% (Avon Woods Nature Preserve) to 100% (Buttercup Valley Nature Preserve). Although the Landsat-8 NDVI change technique generally worked well for detection and mapping of L. maackii in the five urban green spaces of Cincinnati, setting the threshold at 0.035 is, unsurprisingly, imperfect. The omission/commission errors in Table 2 demonstrate errors with the threshold training data. Further research with a separate validation dataset should be collected to determine if the 0.035 threshold is valid at the same sites in different years, and in the same region at different sites.
NDVI difference image of November minus January displaying in white L. maakii (a) Bender Mountain Nature Preserve; (b) Buttercup Valley Nature Preserve and Parkers Wood; (c) Undeveloped wooded section of Spring Grove Cemetery; (d) Avon Woods Nature Preserve.
Errors of commission and omission resulting from application of our proposed L. maackii detection threshold. The threshold had a high average accuracy, but limited ability to detect presence in some sites (e.g., Parker Woods and Avon Woods).
The omission errors in Avon Woods Nature Preserve could correspond to the large patches of evergreens such as Hedera helix (English ivy) and Euonymus fortunei (winter creeper) growing below the L. maackii canopy. This would reduce the November-to-January change in the NIR reflectance signal. English ivy and winter creeper stay green all year creating a low NDVI difference value. In parts of Spring Grove Cemetery and Buttercup Valley/Parker Woods, a commission error could have been caused by green grasses. Grass is likely to stay green longer than the native forest vegetation displaying a similar spectral signature. This suggests that the NDVI change technique must be limited to forested urban green spaces, perhaps by creating vector files of forested boundaries from land-use/land-cover maps before applying the NDVI change technique to Landsat-8 and similar imagery. Additionally, finding satellite imagery without clouds in the region of interest is important and might be variably hard to obtain.
The variability of leaf fall from year to year and within different geographic regions could complicate choosing satellite imagery dates for this change detection method, a limitation also described in Wilfong et al. (2009). It is important to choose an image that has L. maackii with green foliage and the native vegetation without foliage (e.g., the November image in this study). The native vegetation leaf-off time changes depending on the temperature of certain years, so it may require selecting an earlier or later time to obtain a useable image for this detection method. Further research also needs to verify if this proposed threshold holds true in different geographic regions and different appropriate image times.
Our overall mapping accuracy was 82%, which is within the range of accuracy Shouse et al. (2013) found using similar Landsat-5 (30m) imagery. When using this NDVI difference L. maackii detection method with 30m resolution imagery, the L. maackii patch should be at least 3 × 3 pixels or 8,100 m2, otherwise the species likely will not be detected. It makes sense to use this technique in heavily invaded areas (> 8,100m2). To detect smaller patch sizes (< 8,100m2) or early L. maackii detection, the spatial resolution of the imagery needs to be < 20m, although high spatial resolution imagery are often harder to obtain. This study was performed in proprietary ArcGIS software (Environmental Systems Research Institute, Inc., 2010, Redlands, CA) but the calculations are simple and could have easily been run in freeware like QGIS (QGIS Development Team, 2019, Open Source Geospatial Foundation Project-Bonn). Introductory technical skills in spatial data and GIS software is necessary for performing this analysis, an increasingly common skill for biologists.
Meehania cordata. USDA-NRCS PLANTS Database. Britton, N.L. and A. Brown. 1913. An Illustrated Flora of the Northern United States, Canada and the British Possessions. New York, NY: Charles Scribner’s Sons.
Acknowledgements
We would like to thank Tim Sisson of the Western Wildlife Coridor, David Gressley of Spring Grove Cemetery and Arboretum, Skip Meinhardt of Northside Greenspace Inc, and Jim Godby of the Cincinnati Parks for providing advice, information, and support which made this research possible.
This open access article is distributed under the terms of the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0) and is freely available online at:http://jhr.uwpress.org
