- Email: [email protected]
Rapid assessment of juniper distribution in prairie landscapes of the northern Great Plains
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
Contents lists available at ScienceDirect
Int J Appl Earth Obs Geoinformation journal homepage: www.elsevier.com/locate/jag
Rapid assessment of juniper distribution in prairie landscapes of the northern Great Plains
T
Kyle D. Kaskiea, Michael C. Wimberlyb, , Peter J. Baumana ⁎
a b
Department of Natural Resource Management, South Dakota State University, Brookings, SD, 57007, USA Department of Geography and Environmental Sustainability, University of Oklahoma, Norman, OK, 73019, USA
ARTICLE INFO
ABSTRACT
Keywords: Spectral mixture analysis Matched filtering Landsat Accuracy assessment Tree invasion Eastern redcedar
Woody plant species including eastern redcedar (Juniperus virginiana) and Rocky Mountain juniper (Juniperus scopulorum) are expanding throughout the prairie ecosystems of the Great Plains because of fire suppression, land management practices, and climate change. Juniper encroachment threatens native grasslands by altering soil characteristics, limiting herbaceous biomass, hindering native plant regeneration, and reducing rangeland productivity. Existing land cover products do not effectively characterize the distribution of juniper over a range of densities, making it difficult to assess the scale of the problem. We evaluated a method for rapid classification and mapping of juniper using matched filtering with Landsat 8 snow-covered and snow-free winter imagery (January-March), and snow-free spring imagery (April–June) for 2015–2016. We used data from two path/rows (29/30 and 30/30) in southeastern South Dakota and northeastern Nebraska (approximately 23,000 km2). In both path/rows, we found that snow-covered winter images increased contrast between juniper and the image background and resulted in the highest overall classification accuracies of 94.5% and 88.9% for juniper densities above 15%, compared to 91.4% and 85.7% for snow-free winter imagery and 57.8% and 74.1% for growing season imagery. Using winter imagery, we successfully captured pixels containing juniper density above 50% with ≥90% detection probability. However, the true positive rate dropped to less than 50% once juniper density fell below 20%. We identified 84 791 ha within the study area occupied by juniper (3.6% of the total area), including 27 504 ha in deciduous forests (33% of deciduous forest area) and 38 738 ha in grasslands (6% of grassland area).
1. Introduction The Great Plains has been recognized as one of North America’s most endangered ecosystems (Samson and Knopf, 1996). A significant factor driving loss of native ecosystems has been the expansion of agriculture and the resulting conversion of grassland to cultivated cropland (Samson et al., 2004; Wright and Wimberly, 2013; Wimberly et al., 2017). However, the encroachment of woody plants, particularly eastern redcedar (Juniperus virginiana) and Rocky Mountain juniper (Juniperus scopulorum), is also contributing to the widespread degradation and loss of prairie ecosystems in the Great Plains. Historically, the geographic extent of these species (collectively referred to as “juniper” hereafter) was greatly restricted by frequent fires, but in the modern landscape juniper has expanded its range by encroaching into grasslands (Briggs et al., 2002a; Twidwell et al., 2013). A long history of fire suppression and other land use practices have contributed to increases in juniper cover throughout the Great Plains (Briggs et al.,
⁎
2002a). In addition, environmental fluctuations, particularly droughts, can give junipers a competitive advantage over other species and facilitate their expansion (Volder et al., 2013). Juniper encroachment has already overwhelmed much of the southern Great Plains (Norris et al., 2001; Starks et al., 2014) and is spreading rapidly in the North (Meneguzzo and Liknes, 2015; Pierce and Reich, 2010). Woody plant encroachment alters native grasslands by changing soil characteristics, limiting herbaceous biomass, hindering regeneration of native grassland species, and reducing rangeland forage production (Briggs et al., 2002b; Gehring and Bragg, 1992; McKinley and Blair, 2008). Juniper encroachment also results in degradation of habitat quality for grassland birds and other native wildlife species (Coppedge et al., 2004). Management practices such as mechanical removal, herbicide application, and prescribed burning are commonly used to reduce tree cover and restore grassland communities (Wilson and Schmidt, 1990). However, these management activities can be time consuming and costly, and their effectiveness generally decreases as
Corresponding author. E-mail address: [email protected] (M.C. Wimberly).
https://doi.org/10.1016/j.jag.2019.101946 Received 17 June 2019; Received in revised form 13 August 2019; Accepted 16 August 2019 Available online 26 August 2019 0303-2434/ © 2019 Elsevier B.V. All rights reserved.
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
tree size and stand density increase (Bidwell et al., 2002; Buehring et al., 1971; Ortmann et al., 1998). Detailed and accurate maps of juniper distribution can aid in determining the extent of juniper cover, assessing the ecological impacts of juniper encroachment, and developing strategies for prioritizing juniper management across large regions encompassing thousands to millions of hectares. There are multiple sources of data on juniper extent and dynamics, each with particular advantages and limitations. The Forest Inventory and Analysis (FIA) program of the USDA Forest Service collects field data on a sample of one-acre forestland plots within a state, allowing computation of yearly forest type estimates and providing a complete statewide inventory at five year intervals. Meneguzzo and Liknes (2015) used these data to show that over a seven-year time span, (2005–2012) the juniper forest type increased by 116 000 ha in the central United States, resulting in an annual loss of 16 500 ha of nonforestland to juniper encroachment between 2007 and 2012. FIA data have also been combined with Moderate Resolution Imaging Spectroradiometer (MODIS) 250 m imagery to interpolate and produce live volume and density maps of juniper (Meneguzzo et al., 2008; Simonsen et al., 2015). Although these coarse-grained maps and area estimates are suitable for general assessments of juniper distribution over large areas, they do not precisely depict all juniper locations in the landscape. Accurate higher-resolution maps are needed to make large-area assessments and inform site-specific management decisions. Previous studies have used a variety of remote sensing data, including very high spatial resolution (VHSR) aerial imagery (Anderson and Cobb, 2004; Poznanovic et al., 2014), hyperspectral imagery (Wylie et al., 2000), and fusion of Landsat Thematic Mapper (TM) imagery with Light Detection and Ranging (LiDAR) data (Sankey et al., 2010). A more recent study has shown the potential for mapping eastern redcedar at a large scale with the combination of L-band Synthetic Aperture Radar (SAR) data and Landsat Thematic Mapper/Enhanced Thematic Mapper Plus (TM/ETM +) imagery (Wang et al., 2017). Although these studies have provided useful results, many of the data sources have limited spatial or temporal coverage, making these approaches difficult to apply to other study areas. Landsat imagery is a freely accessible data source that allows for large-scale mapping and continuous monitoring of woody plant encroachment (Sankey and Germino, 2008; Venter et al., 2018; Wang et al., 2017; Yang et al., RSE 2012). Since 1984, the Landsat program has provided 30-m multispectral images at a 16-day temporal resolution (Wulder et al., 2006). The Landsat 8 mission, launched in 2013, features a global acquisition strategy, has improved radiometric resolution compared to previous Landsat sensors, and eliminates the data gaps associated with the scan-line corrector failure on Landsat 7. One limitation of Landsat is that the pixel size is larger than most individual trees, so that pixels often contain a mixture of trees and herbaceous grassland vegetation. For this reason, sub-pixel methods such as spectral mixture analysis have been utilized for classifying juniper encroachment with Landsat imagery (Sankey and Germino, 2008). To generate juniper maps that are useful for management purposes, there is a need for replicable classification methods that are straightforward to implement and use readily available open source data to support high-accuracy mapping. To address this need, we investigated the classification of juniper in a prairie landscape using a linear spectral unmixing method with Landsat 8 satellite imagery. Our objectives were to: 1) evaluate the potential for juniper detection using partial unmixing techniques with Landsat 8 OLI (Operational Land Imager) imagery; 2) compare classification accuracies for imagery from different seasons with distinctive spectral characteristics (growing season, non-growing season with consistent snow coverage, and non-growing season imagery with no snow coverage); 3) assess classification accuracy for varying levels of juniper density; and 4) compare patterns of juniper distributions with generic land cover classes from the National Land Cover Database to determine the new information gained from the juniper map.
2. Material and Methods 2.1. Study Area The study area covered 14 contiguous counties bordering the Missouri River, including nine counties in southeastern South Dakota and five counties in northeastern Nebraska (Fig. 1). This area has a Köppen climate classification of humid continental (Dfa) (Kottek et al., 2006), with an annual mean temperature varying between 6–11 °C and an annual precipitation varying between 498–796 mm from 1981 through 2010. Steeply sloped drainages disrupt a flat to rolling topography with land cover comprised mostly of cropland (48%) and herbaceous grasslands (39%). The primary land uses include the agricultural production of corn, soybeans, and wheat as well as cattle grazing. Common native grassland vegetation consists of mixed-grass prairie species such as little bluestem (Schizachyrium scoparium), big bluestem (Andropogon gerardii), western wheatgrass (Pascopyrum smithii), sideoats grama (Bouteloua curtipendula), and green needlegrass (Nassella viridula). Woodlands are primarily found near drainages and riparian lowlands with the exception of small groves scattered across the prairie uplands. Common deciduous tree species include the plains cottonwood (Populus deltoids), bur oak (Quercus macrocarpa), green ash (Fraxinus pennsylvanica) and American elm (Ulmus americana). Juniper species such as Rocky mountain juniper (Juniperus scopulorum) and eastern redcedar (Juniperus virginianai) are also common (Barker and Whitman, 1988). 2.2. Data sources We obtained Landsat 8 Operational Land Imager (OLI) level-2 surface reflectance data for path/row 29/30 and 30/30 (Fig. 1) containing minimal cloud cover (< 10%) through the U.S. Geological Survey (USGS) EarthExplorer online tool (USGS Earth Explorer, 2017). These data were generated using the Landsat Surface Reflectance Code (LaSRC). We inspected each image and its associated pixel quality band and selected three images with minimal cloud cover that covered the study area for each path/row. We chose images from different seasons with distinctive spectral characteristics, including one image from the growing season (April through July) with mostly green vegetation, and two images from the non-growing season (January through March). For the non-growing season, we selected one image with uniform snow coverage and one image with no snow coverage that was predominantly senesced vegetation and soil. Snow cover has been shown to reduce spectral variability among ground surface targets when using Landsat sensor data while also increasing contrast between white snow and forest elements foliage, bark, and shadows above the snow cover (Wolter et al., 2008, 2012). Table 1 shows the final six images chosen for analyses. Once we identified suitable image dates, we extracted surface reflectance data for six bands: band 2 (0.435–0.451 μm, blue), band 3 (0.533–0.590 μm, green), band 4 (0.636–0.673 μm, red), band 5 (0.851–.0879 μm, near infrared), band 6 (1.566–1.651 μm, shortwave infrared), and band 7 (2.107–2.294 μm, shortwave infrared). To reduce spectral misclassification, we constructed a mask for each image that excluded any permanent water body features (i.e. streams, rivers, ponds and lakes). We identified water bodies using the USGS National Hydrography Dataset (NHD), obtained through Geospatial Data Gateway (NRCS, 2017). 2.3. Juniper classification We performed juniper classification using a matched filtering approach with ENVI version 5.4 (Exelis Visual Information Solutions, Boulder, Colorado). Matched filtering is a partial unmixing procedure that incorporates user-defined endmembers to maximize the response 2
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 1. Study area composed of nine counties in southeastern South Dakota and five counties in northeastern Nebraska. Landsat 8 path/rows: 30/30 (a) and 29/30 (b) cover the 14 contiguous counties. Points indicate the locations of visually interpreted juniper presence/absence pixels for validation along with randomly generated and opportunistically chosen field investigation sites.
of known spectral signatures while suppressing unknown background signatures. The required inputs for matched filtering included the spectral signature of a pure endmember, the selection of bands to use with the algorithm, and the selection of a threshold to classify pixels as juniper on non-juniper. Matched filtering requires only the input of the desired endmember or cover type that the user wishes to identify. This approach differs from conventional linear spectral mixture analysis, which requires an input of all known endmembers. Previous partial unmixing research has indicated that the mean of manually selected endmembers containing a high percentage of target cover outperformed extreme or variant n-dimensional visualizer (ND-V) endmember pixels and the mean of all NDV endmember pixels (Sankey and Glenn, 2011). Therefore, we selected ten pixels that were predominantly juniper within each path/row (29/ 30 and 30/30). This process allowed the average spectral signature to be obtained, which we used as the endmember input for the matched filtering analyses.
Previous studies and preliminary observations suggested that a combination of bands 2–5 and band 7 allowed for the best spectral separation between juniper and background materials (Vikhamar and Solberg, 2003). We therefore used these bands to generate a matched filtering image containing values that represented the relative degree of match for each pixel. A value of one signified a perfect match while values closer to zero indicated background or non-target materials. Once we completed the matched filtering analysis, we developed a binary juniper classification map by designating a threshold for the matched filtering values to discriminate juniper from non-juniper pixels. We determined a threshold for each image by sampling a group of user-defined pixels that contained a high percentage (> 50%) of juniper cover. We then computed the mean matched filtering value and standard deviation for the sampled pixels and assigned two negative standard deviations from the mean as the threshold. This threshold was applied to produce a final classified map that contained pixels representing juniper and non-juniper.
Table 1 Landsat 8 OLI Surface Reflectance Level-2 imagery used for the classification of juniper. Product Identifier
Date
Path/row
Cloud cover (%)
Snow cover (%)
LC08_L1TP_029030_20150107_20170302_01_T2 LC08_L1TP_029030_20150328_20170227_01_T1 LC08_L1TP_029030_20160602_20170223_01_T1 LC08_L1TP_030030_20160101_20180131_01_T1 LC08_L1TP_030030_20160305_20170224_01_T1 LC08_L1TP_030030_20160524_20180131_01_T1
2015/01/07 2015/03/28 2016/06/02 2016/01/01 2016/03/05 2016/05/24
29/30 29/30 29/30 30/30 30/30 30/30
1.55 1.11 7.41 0.05 0.45 0.00
94.75 0.00 0.00 97.91 0.00 0.00
3
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 2. Qualitative assessment of juniper classification maps produced by Landsat 8 imagery during (a) the non-growing season with no snow coverage, (b) the nongrowing season with consistent snow coverage, and (c) the growing season (c). Pixels classified as juniper are displayed over 2016 NAIP imagery.
2.4. Accuracy assessment
characteristics (e.g. land use, species composition). We also took additional photos at opportunistic stops. We collected data for 252 sites and obtained 1271 photos distributed throughout our study area (Fig. 1).
To assess accuracy, we used a random stratified sampling design that allowed us to determine the classification accuracy over a range of juniper densities. We specified four strata, which included closed canopy woodlands, buffered closed canopy woodlands, planted shelterbelts, and other non-woodland areas. We digitized the closed canopy woodlands and planted shelterbelts in ArcGIS 10.5 (ESRI Inc., Redlands CA) following the guidelines presented in Bauman et al. (2016) using very high spatial resolution (VHSR) 60 cm National Agricultural Imagery Program (NAIP), 2014 and 2016 aerial imagery. In addition to obtaining samples of dense woodland cover, we also sampled open canopied woodlands and sparse trees. To do so, we placed a 90-m buffer around the digitized closed canopy stratum using ArcGIS. Based on visual observations, we determined that this distance encompassed many areas with low densities of regenerating juniper. Once we defined all the strata within the study area, we generated random points using ArcGIS. We referenced each random point to a Landsat pixel by converting it to a 30 × 30 m polygon and snapping it to the Landsat 8 grid. Then, we characterized each polygon using a combination of VHSR imagery, which included NAIP 2016 and other sources of winter imagery accessed through Google Earth from 2013 to 2017. Data collected for each sample included the land cover type, juniper presence/absence, and the percent of juniper cover within the 30 × 30 m polygon. We delineated a total of 1,643 juniper presence and 2,273 juniper absence points for accuracy assessment covering a range of juniper densities (Fig. 1). We conducted field investigations during October 2017 to obtain ground reference data and gain a better understanding of juniper distribution over the study area. Random locations were selected along road corridors and used to assess 0.5 km driven transects. Each transect was divided into three stops, and each stop was divided into a left and right side (six stops per transect). At each stop, we captured a photo with a GPS enabled digital camera and recorded vegetation and land cover
2.5. Comparison with land cover The National Land Cover Database (NLCD) provides a consistent, national-level map of generic land cover classes at a 30-m spatial resolution that is updated on a regular basis (Homer et al., 2015). To determine whether the juniper map developed in this study provided novel information not already available through the NLCD, and to provide a broader context for interpreting the juniper patterns, we conducted a cross-tabulation of juniper presence/absence with the 2016 NLCD. We assessed the degree to which the juniper maps corresponded with evergreen and mixed forest mapped in the NLCD, and the distribution of juniper across the various NLCD classes. 3. Results We produced a total of six Landsat 8 juniper classification maps under three image conditions (consistent snow coverage during the non-growing season, no snow coverage during the non-growing season, and growing season) for two path/rows: 29/30 and 30/30. A visual assessment using VHSR imagery and field photographs found a reasonably accurate representation of juniper for all images in the nongrowing season (Fig. 2a, b). However, misclassification of non-juniper (i.e. cultivated fields and wetlands) as juniper was observed more frequently in images containing no snow coverage (Fig. 2a) than it was in images containing consistent snow coverage (Fig. 3b). Maps based on growing season imagery appeared to have a higher misclassification of both juniper and non-juniper sites (Fig. 2c), which was confirmed by the quantitative accuracy assessments. 4
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 3. Pixel-level accuracy assessment for the classification of juniper by density with non-growing season imagery. Path/row: (a) 29/30 and (b) 30/30.
3.1. Accuracy assessment
values of 92.1% and 84.7% for path/rows 29/30 and 30/30 compared to 83.2% and 76.3% for snow-covered non-growing season imagery and 70.9% and 70.0% for non-growing season imagery without snow cover. However, both growing season images had much lower user accuracies (the probability that a pixel mapped as juniper actually represents juniper in the reference dataset) of 38.7% and 67.4% compared to 96.3% and 99.2% for snow-covered non-growing season and 96.8% and 98.3% for non-growing season imagery without snow cover. Because of the low user accuracies for the maps based on growingseason imagery, we focused the next phase of the analysis on the maps
The classified maps based on imagery with consistent snow coverage during the non-growing season had the highest accuracy statistics for both path/row combinations (Table 2). These maps had overall accuracy (OA) of 94.5% and 88.9% for path/rows 29/30 and 30/30 compared to 91.4% and 85.7% for non-snow coverage imagery during the non-growing season and 57.8% and 74.1% for the growing season. Producer accuracies (the probability that a reference pixel of juniper is correctly classified) were highest for the growing season imagery with
Table 2 Accuracy assessment for six Landsat 8 scenes based on classified juniper presence and absence pixels. Validation points were considered to have juniper present if they had a juniper cover greater than 15%. Shown underlined, overall accuracy (OA) represents the total classification accuracy for both juniper presence and juniper absence pixels. Path/Row
Condition
Classified pixel data
Reference pixel data
User Accuracy (UA)
Juniper absence
Juniper presence
Row totals
29/30
Snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
887 11 898 0.9878
57 283 340 0.8324
944 294 1238
0.9396 0.9626
29/30
Non-snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
890 8 898 0.9911
99 241 340 0.7088
989 249 1238
0.8999 0.9679
29/30
Growing
Juniper absence Juniper presence Column totals Producer accuracy (PA)
403 495 898 0.4488
27 313 340 0.9206
430 808 1238
30/30
Snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
978 5 983 0.9949
198 639 837 0.7634
1176 644 1820
30/30
Non-snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
973 10 983 0.9898
251 586 837 0.7001
1224 596 1820
Juniper absence Juniper presence Column totals Producer accuracy (PA)
640 343 983 0.6511
128 709 837 0.8471
768 1052 1820
30/30
Growing
5
0.9101
0.8500 0.9372 0.3874 0.6847 0.8316 0.9922 0.8792 0.7949 0.9832 0.8450 0.8333 0.6740 0.7491
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 4. Examples of juniper density at (a) 1–10, (b) 10–20, (c) 20–30, (d) 30–40, (e) 40–50, (f) 50–60, (g) 60–70, (h) 70–80, (i) 80–90, and (j) 90–100%. 30 × 30 m polygons are overlaid on NAIP and Google Earth imagery.
based on non-growing season imagery. The true positive rate (calculated as the percent of juniper validation points that were correctly classified) of matched filtering classification of juniper for non-growing season imagery was assessed at ten juniper density levels (Figs. 3 and 4). The true positive rate of juniper for both image conditions in the non-growing season was 90% or higher (Fig. 3) for pixels containing juniper density above 50% (Fig. 4f–j). When juniper density was below 50%, the true positive rate varied more between each path/row and image condition (snow versus no snow cover) than for pixels containing higher densities of juniper. Imagery with consistent snow coverage produced higher true positive rates at lower levels of juniper coverage than images with no snow coverage (Fig. 3). The true positive rate dropped below 50% for non-snow covered images when juniper densities were at 20–30% (Fig. 4c; path/row: 29/30 and 30/30; 47% and 40%) and for snow-covered images at 10–20% (Fig. 4b; 48% and 37%). For pixels with the lowest level of juniper cover ranging from 1 to 10% (Fig. 4a), true positive rates were extremely low, with a probability of detection less than 13% for all images.
4. Discussion 4.1. Map accuracy Our visual and quantitative accuracy assessments demonstrate that matched filtering of Landsat 8 OLI imagery with consistent snow coverage during the non-growing season is an effective technique for mapping landscape-level patterns of juniper across large areas in the northern Great Plains. Practical advantages are that this technique requires spectral data from a relatively small number of juniper-dominated pixels to characterize the juniper endmember and can be implemented with a single Landsat scene. If a suitable cloud-free, snowcovered imagery is available for the location and time of interest, it is possible to rapidly and efficiently generate juniper distribution maps. Using the matched filtering approach, we were able to obtain accuracies comparable to other juniper-mapping studies that used multiple data sources (Sankey et al., 2010; Wang et al., 2017) or hyperspectral data (Wylie et al., 2000). When images from the non-growing season were used, the resulting maps captured pixels containing juniper density above 50% with a 90% and greater detection probability, and we also obtained a true positive rate of 50% or greater for pixels containing juniper density above 15%. Wang et al. (2017) similarly achieved a 90% detection probability for pixels containing eastern redcedar (Juniperus virginiana) density above 60% and a 30% detection probability when pixel densities where between 10 to 20% using a combination of multi-temporal Landsat imagery and L-band SAR imagery. We were able to obtain comparable classification accuracies for a single-date map based on a single Landsat 8 image. The Landsat-based juniper classification maps produced higher overall classification accuracies for images during the non-growing season compared to images obtained during the growing season. This difference can be attributed to smaller differences in spectral signatures between juniper and other types of vegetation and woodlands during the growing season compared to the winter months, when juniper was the only green vegetation (Wang et al., 2017). We were also able to obtain a better spectral separation of juniper and non-juniper when images containing consistent snow coverage were used compared to images containing no snow coverage during the non- growing season. The snow-covered pixels contained higher reflectance values in the visible wavelengths and near-infrared wavelengths (Dozier and Painter, 2004; Vikhamar and Solberg, 2003). This bright background provided a strong contrast to juniper, which had very low reflectance in the visible
3.2. Juniper patterns and land cover associations A final juniper map was created by mosaicking the juniper classification maps produced from snow-covered imagery for path/rows 29/30 and 30/30. This map, shown in Fig. 5a, identified areas of planted juniper while also detecting areas of naturally-regenerated juniper (Fig. 5b, c). Examination of the final map also showed that the matched filtering technique with snow-covered imagery could effectively distinguish juniper cover from surrounding deciduous woodlands (Fig. 5d, e). The juniper map identified 84 791 ha occupied by juniper (3.6% of the total area). Most of the evergreen forest (89%) and mixed forest (61%) classes from the NLCD were occupied by juniper in our map (Table 3). In addition, 33% of deciduous forest, 17% of woody wetlands, 10% of shrublands, and 6% of grasslands were occupied by juniper. However, only 11% of the juniper area was classified as either evergreen or mixed forest in the NLCD. In addition, 32% of the juniper area (27 504 ha) was classified as deciduous forest and 46% of the juniper area (38 738 ha) was classified as grassland. Only small percentages of the juniper area (< 3%) were classified as other NLCD classes. Overall, the NLCD evergreen and mixed forest classes substantially underrepresented the extent of juniper, which is the most widespread evergreen tree species within the study area. 6
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 5. (a) Final juniper classification map. (b, c) Close-up maps from region 2 showing juniper spread from planted shelterbelts. (d,e) Close-up maps from region 1 showing the discrimination of juniper from deciduous trees. A Google Earth high-resolution image dated 4/11/2016 shows junipers as green while leaf-off deciduous trees are gray-brown. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
wavelengths and near-infrared wavelengths compared to the snowcovered pixels. The snow cover aided in the classification of juniper by reducing variation in the background non-juniper matrix, permitting our material of interest (i.e., juniper) to be the primary driver of the scene covariance (Boardman et al., 1995). However, a limitation to the use of snow-covered images is that they may not be available for every year and location of interest, and they will become progressively more difficult to obtain in warmer locations where snowfall is less common. The pixel-level accuracy assessments showed a strong trend associated with juniper density, where probability of correctly classifying a pixel as containing juniper increased with juniper density. This
observation was also noted in the recent study of Wang et al. (2017), who suspected that incorrectly classified low-density juniper pixels were influenced by juniper height and the omission of woodlands that did not meet their definition of a forest. In our study the loss of accuracy for low-density juniper, particularly at levels of cover less than 20%, can be attributed to the weaker spectral signatures of pixels containing only a small fraction of juniper cover. Low-density juniper pixels contain more non-juniper vegetation in a heterogeneous background, whereas pixels with higher juniper density have less background vegetation (Briggs et al., 2002b; Gehring and Bragg, 1992). These differences influence the partial spectral unmixing classification because
Table 3 Cross-tabulated areas from the overlay of the classified juniper map with land cover classes the 2016 National Land Cover Database (NLCD). NLCD Class
Non-Juniper (ha)
Juniper (ha)
Total (ha)
% NLCD Classified as Juniper
% Juniper in NLCD Class
Water Developed (Open) Developed (Low) Developed (Medium) Developed (High) Barren Deciduous Forest Conifer Forest Mixed Forest Shrub Grassland Pasture/Hay Crop Woody Wetland Emergent Wetland Total
77 174 65 508 12 221 3 259 935 1 869 54 611 995 489 5 315 645 345 201 900 1 169 616 10 773 38 830 2 288 840
365 965 297 37 4 19 27 504 8 211 773 606 38 738 1 818 2 357 2 216 880 84 791
77 540 66 473 12 518 3 296 939 1 888 82 115 9 206 1 262 5 920 684 083 203 718 1 171 973 12 988 39 710 2 373 630
0.47 1.45 2.37 1.14 0.43 1.01 33.49 89.19 61.28 10.23 5.66 0.89 0.20 17.06 2.22
0.43 1.14 0.35 0.04 0.00 0.02 32.44 9.68 0.91 0.71 45.69 2.14 2.78 2.61 1.04
7
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
non-juniper spectral signatures mask the spectral signatures of junipers within the pixel. If the classification threshold was decreased, the probability of positively identifying lower density pixels would increase, but at the expense of increasing the classification of more nonjuniper sites as juniper (i.e., false positives). The low classification accuracy for low-density juniper pixels is a limitation of our mapping approach. Classifying these low-density juniper pixels can be particularly important because these are the areas where juniper removal via fire or mechanical cutting is most cost-effective and where grassland vegetation is still present and there is the greatest potential for prairie restoration. Because of the constraints imposed by Landsat’s 30-m resolution, additional data sources and methods will be needed to reliably identify low-density juniper. One possibility is to use very high-resolution aerial or satellite imagery to identify and map individual juniper trees (Falkowski et al., 2017; Ozdemir, 2008; Poznanovic et al., 2014). Another option could be to predict areas that are likely to support low densities of juniper regeneration based on the locations of juniper seed sources and environmental settings conducive to juniper establishment (Greene and Knox, 2014; Weisberg et al., 2007).
communities (Knapp et al., 2008; Ratajczak et al., 2012; Van Auken, 2009). Accurate and timely data on the distribution of woody plant encroachment is essential for assessing the economic and ecological impacts of woody encroachment (Anadón et al., 2014; Zavaleta, 2000) and will assist in establishing and implementing the appropriate management measures (Bidwell et al., 2002; Ortmann et al., 1998; Smith, 2011). Juniper species are the main woody plant species threatening grassland habitats in our region (Engle et al., 2008), and we were able to develop maps that accurately depict the distribution of juniper by using partial unmixing techniques with Landsat 8 OLI imagery. This workflow provides a straightforward means for mapping juniper distribution over a large scale, allowing managers to quantify the regional impact of juniper expansion and identify sites where further encroachment is likely in the future. Acknowledgements We thank Izaya Numata and Justin Davis for advice on remote sensing and analytical methods. Joshua Leffler provided helpful comments on an earlier draft of the manuscript. This research was supported by the Geospatial Sciences Center of Excellence at South Dakota State University.
4.2. Applications of the juniper distribution map The methods developed in this study allows for the mapping of juniper distribution across large regions and can facilitate regional reassessment of management polices (Roberts et al., 2018). Juniper maps can also aid in monitoring to determine where juniper has encroached into environmentally sensitive areas, supporting management decisions that can help prevent of future grassland loss (Bauman et al., 2016; Wimberly et al., 2018). The juniper maps developed through this research will also establish a baseline for future studies. Although this study focused on only two Landsat path/rows and a single date, the same partial unmixing techniques could be applied across larger extents and over multiple time periods. The use of remote sensing to address the dynamic encroachment of juniper at a large scale has only recently been investigated in the southern Great Plains (Wang et al., 2018), but is yet to be explored in the northern Great Plains. With a better understanding of the drivers associated with juniper encroachment, current ecological knowledge (Briggs et al., 2002b; Pierce and Reich, 2010) and the development of accurate juniper maps, we can manage encroaching juniper species more efficiently by targeting proactive management strategies and restoration activities (Greene and Knox, 2014). The juniper map yielded new information that complements the data provided by the NLCD, a standard national-level land cover data product. The NLCD has been developed to provide accurate mapping of generic land cover classes at regional to national levels (Homer et al., 2015) but is not optimal for capturing sub-pixel land-cover features such as juniper encroachment. Although the NLCD forest classes are specified to identify areas with more than 20% trees, the total area of evergreen and mixed forests was much smaller than the area occupied by juniper in our study area. Most of the juniper areas were located in grasslands with partial juniper cover or in deciduous forest stands with a juniper component. Our juniper map provides unique regionally-relevant information that complements the NLCD, and we have demonstrated that combining these two data sources can provide valuable information about the extent and pattern of juniper encroachment. For example, we estimated that 6% of the grasslands in the study area, encompassing 38 783 ha, are currently being impacted by juniper encroachment, and the juniper map can be used to find the locations of these areas. In contrast, interpreting the NLCD conifer and mixed forest classes as a measurement of juniper distribution would greatly underestimate the impacted area. Although this study focused primarily on mapping juniper encroachment in a portion of the northern Great Plains, woody plant encroachment is also threatening other North American grassland
References Anadón, J.D., Sala, O.E., Turner, B.L., Bennett, E.M., 2014. Effect of woody-plant encroachment on livestock production in North and South America. Proc. Natl. Acad. Sci. U.S.A. 111, 12948–12953. Anderson, J.J., Cobb, N.S., 2004. Tree cover discrimination in panchromatic aerial imagery of pinyon-juniper woodlands. Photogramm. Eng. Remote Sensing 70, 1063–1068. Barker, W.T., Whitman, W.C., 1988. Vegetation of the northern Great Plains. Rangelands 10, 266–272. Bauman, P., Carlson, B., Butler, T., 2016. Quantifying Undisturbed (Native) Lands in Eastern South Dakota: 2013. South Dakota State University Extension, Brookings, SD. Bidwell, T.G., Weir, J.R., Engle, D.M., 2002. Eastern Redcedar Control and Management: Best Management Practices to Restore Oklahoma’s Ecosystems. Oklahoma Cooperative Extension Service Fact Sheet NREM-2876. Briggs, J.M., Knapp, A.K., Brock, B.L., 2002a. Expansion of woody plants in tallgrass prairie: a fifteen-year study of fire and fire-grazing interactions. Am. Midl. Nat. 147, 287–294. Briggs, J.M., Hoch, G.A., Johnson, L.C., 2002b. Assessing the rate, mechanisms, and consequences of the conversion of tallgrass prairie to Juniperus virginiana forest. Ecosystems 5, 578–586. Boardman, J.W., Kruse, F.A., Green, R.O., 1995. Mapping target signatures via partial unmixing of AVIRIS data. In: Proceedings of the Fifth JPL Airborne Earth Science Workshop. Pasadena, CA. pp. 95–101. Buehring, N., Santelmann, P.W., Elwell, H.M., 1971. Responses of eastern red cedar to control procedures. J. Range Manag. 24, 378–382. Coppedge, B.R., Engle, D.M., Masters, R.E., Gregory, M.S., 2004. Predicting juniper encroachment and CRP effects on avian community dynamics in southern mixed-grass prairie. USA. Biol. Conserv 115, 431–441. Dozier, J., Painter, T.H., 2004. Multispectral and hyperspectral remote sensing of alpine snow properties. Annu. Rev. Earth Planet. Sci. 32, 465–494. Engle, D.M., Coppedge, B.R., Fuhlendorf, S.D., 2008. From the dust bowl to the green glacier: human activity and environmental change in Great plains grasslands. In: van Auken, O. (Ed.), Western North American Juniperus Communities. Springer, New York, pp. 253–271. Falkowski, M.J., Evens, J.S., Naugle, D.E., Hagen, C.A., Carleton, S.A., Maestas, J.D., Khalyani, A.H., Poznanovic, A.J., Lawrence, A.J., 2017. Mapping tree canopy cover in support of proactive prairie grouse conservation in western North America. Rangel. Ecol. Manag. 70, 15–24. Gehring, J.L., Bragg, T.B., 1992. Changes in prairie vegetation under eastern red cedar (Juniperus virginiana L.) in an eastern Nebraska bluestem prairie. Am. Midl. Nat. 128, 209–217. Greene, S.L., Knox, J.C., 2014. Coupling legacy geomorphic surface facies to riparian vegetation: assessing red cedar invasion along the Missouri River downstream of Gavins Point dam, South Dakota. Geomorphology 204, 277–286. Homer, C., Dewitz, J., Yang, L., Jin, S., Danielson, P., Xian, G., Coulston, J., Herold, N., Wickham, J., Megown, K., 2015. Completion of the 2011 National Land Cover Database for the conterminous United States–representing a decade of land cover change information. Photogramm. Eng. Remote Sensing 81, 345–354. Knapp, A.K., Briggs, J.M., Collins, S.L., Archer, S.R., Bret-Harte, M.S., Ewers, B.E., Peters, D.P., Young, D.R., Shaver, G.R., Pendall, E., Cleary, M.B., 2008. Shrub encroachment in North American grasslands: shifts in growth form dominance rapidly alters control of ecosystem carbon inputs. Glob. Chang. Biol. 14, 615–623. Kottek, M., Grieser, J., Beck, C., Rudolf, B., Rubel, F., 2006. World map of the KöppenGeiger climate classification updated. Meteorol. Z. 15, 259–263.
8
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al. McKinley, D.C., Blair, J.M., 2008. Woody plant encroachment by Juniperus virginiana in a mesic native grassland promotes rapid carbon and nitrogen accrual. Ecosystems 11, 454–468. Meneguzzo, D.M., Butler, B.J., Crocker, S.J., Haugen, D.E., Moser, W.K., Perry, C.H., Wilson, B.T., Woodall, C.W., 2008. Nebraska’s forests, 2005. Resour. Bull. NRS-27. U.S. Department of Agriculture Forest Service, Northern Research Station, Newtown Square, PA. Meneguzzo, D.M., Liknes, G.C., 2015. Status and trends of eastern redcedar (Juniperus virginiana) in the central United States: analyses and observations based on Forest Inventory and Analysis data. J. For. 113, 325–334. Norris, M.D., Blair, J.M., Johnson, L.C., 2001. Land cover change in eastern Kansas: litter dynamics of closed-canopy eastern redcedar forests in tallgrass prairie. Can. J. Bot. 79, 214–222. Ortmann, J., Stubbendieck, J., Masters, R.A., Pfeiffer, G.H., Bragg, T.B., 1998. Efficacy and costs of controlling eastern redcedar. J. Range Manag. 51, 158–163. Ozdemir, I., 2008. Estimating stem volume by tree crown area and tree shadow area extracted from pan-sharpened Quickbird imagery in open Crimean juniper forests. Int. J. Remote Sens. 29, 5643–5655. Pierce, A.M., Reich, P.B., 2010. The effects of eastern red cedar (Juniperus virginiana) invasion and removal on a dry bluff prairie ecosystem. Biol. Invasions 12, 241–252. Poznanovic, A.J., Falkowski, M.J., Maclean, A.L., Smith, A., Evans, J.S., 2014. An accuracy assessment of tree detection algorithms in juniper woodlands. Photogramm. Eng. Remote Sensing 80, 627–637. Ratajczak, Z., Nippert, J.B., Collins, S.L., 2012. Woody encroachment decreases diversity across North American grasslands and savannas. Ecology 93, 697–703. Roberts, C.P., Uden, D.R., Allen, C.R., Twidwell, D., 2018. Doublethink and scale mismatch polarize policies for an invasive tree. PLoS One 13, e0189733. Samson, F.B., Knopf, F.L. (Eds.), 1996. Prairie Conservation: Preserving North America’S Most Endangered Ecosystem. Island Press, Washington D.C. Samson, F., Knopf, F., Ostlie, W., 2004. Great Plains ecosystems: past, present, and future. Wildl. Soc. Bull. 32, 6–15. Sankey, T.T., Germino, M.J., 2008. Assessment of juniper encroachment with the use of satellite imagery and geospatial data. Rangel. Ecol. Manag. 61, 412–418. Sankey, T.T., Glenn, N., Ehinger, S., Boehm, A., Hardegree, S., 2010. Characterizing western juniper expansion via a fusion of Landsat 5 Thematic Mapper and lidar data. Rangel. Ecol. Manag. 63, 514–523. Sankey, T., Glenn, N., 2011. Landsat-5 TM and lidar fusion for sub-pixel juniper tree cover estimates in a western rangeland. Photogramm. Eng. Remote Sensing 77, 1241–1248. Simonsen, V.L., Fleishmann, J.E., Whisenhunt, D., Volesky, J.D., Twidwell, D., 2015. Act Now or Pay Later: Evaluating the Cost of Reactive Versus Proactive Eastern Redcedar Management. University of Nebraska-Lincoln Extension Report EC1784. Smith, S., 2011. Eastern Red-cedar: Positives, Negatives and Management. Samuel Roberts Noble Foundation, Ardmore, OK. Starks, P.J., Venuto, B.C., Dugas, W.A., Kiniry, J., 2014. Measurements of canopy interception and transpiration of eastern redcedar grown in open environments. Environ. Nat. Resour. Res. 4, 103–122. Twidwell, D., Rogers, W.E., Fuhlendorf, S.D., Wonkka, C.L., Engle, D.M., Weir, J.R., Kreuter, U.P., Taylor, C.A., 2013. The rising Great Plains fire campaign: citizens’
response to woody plant encroachment. Front. Ecol. Environ. 11, e65–e71. Van Auken, O.W., 2009. Causes and consequences of woody plant encroachment into western North American grasslands. J. Environ. Manage. 90, 2931–2942. Venter, Z.S., Cramer, M.D., Hawkins, H.J., 2018. Drivers of woody plant encroachment over Africa. Nat. Commun. 9, 2272. Vikhamar, D., Solberg, R., 2003. Subpixel mapping of snow cover in forests by optical remote sensing. Remote Sens. Environ. 84, 69–82. Volder, A., Briske, D.D., Tjoelker, M.G., 2013. Climate warming and precipitation redistribution modify tree-grass interactions and tree species establishment in a warmtemperate savanna. Glob. Chang. Biol. 19, 843–857. Wang, J., Xiao, X., Qin, Y., Dong, J., Geissler, G., Zhang, G., Cejda, N., Alikhani, B., Doughty, R.B., 2017. Mapping the dynamics of eastern redcedar encroachment into grasslands during 1984–2010 through PALSAR and time series Landsat images. Remote Sens. Environ. 190, 233–246. Wang, J., Xiao, X., Qin, Y., Doughty, R.B., Dong, J., Zou, Z., 2018. Characterizing the encroachment of juniper forests into sub-humid and semi-arid prairies from 1984 to 2010 using PALSAR and Landsat data. Remote Sens. Environ. 205, 166–179. Weisberg, P.J., Lingua, E., Pillai, R.B., 2007. Spatial patterns of pinyon–juniper woodland expansion in Central Nevada. Rangel. Ecol. Manag. 60, 115–124. Wilson, J., Schmidt, T., 1990. Controlling eastern redcedar on rangelands and pastures. Rangelands 12, 156–158. Wimberly, M.C., Janssen, L.L., Hennessy, D.A., Luri, M., Chowdhury, N.M., Feng, H., 2017. Cropland expansion and grassland loss in the eastern Dakotas: new insights from a farm-level survey. Land Use Policy 63, 160–173. Wimberly, M.C., Narem, D.M., Bauman, P.J., Carlson, B.T., Ahlering, M.A., 2018. Grassland connectivity in fragmented agricultural landscapes of the north-central United States. Biol. Conserv. 217, 121–130. Wolter, P.T., Townsend, P.A., Sturtevant, B.R., Kingdon, C.C., 2008. Remote sensing of the distribution and abundance of host species for spruce budworm in Northern Minnesota and Ontario. Remote Sens. Environ. 112, 3971–3982. Wolter, P.T., Berkley, E.A., Peckham, S.D., Singh, A., Townsend, P.A., 2012. Exploiting tree shadows on snow for estimating forest basal area using Landsat data. Remote Sens. Environ. 121, 69–79. Wright, C.K., Wimberly, M.C., 2013. Recent land use change in the Western Corn Belt threatens grasslands and wetlands. Proc. Natl. Acad. Sci. U.S.A. 110, 4134–4139. Wulder, M.A., Dymond, C.C., White, J.C., Leckie, D.G., Carroll, A.L., 2006. Surveying mountain pine beetle damage of forests: a review of remote sensing opportunities. For. Ecol. Manage. 221, 27–41. Wylie, B.K., Meyer, D.J., Choate, M.J., Vierling, L., Kozak, P.K., Green, R.O., 2000. Mapping woody vegetation and eastern red cedar in the Nebraska Sand Hills using AVIRIS. AVIRIS Airborne Geoscience Workshop. JPL Publication, Pasadena, CA, pp. 00–18. Yang, J., Weisberg, P.J., Bristow, N.A., 2012. Landsat remote sensing approaches for monitoring long-term tree cover dynamics in semi-arid woodlands: comparison of vegetation indices and spectral mixture analysis. Remote Sens. Environ. 119, 62–73. Zavaleta, E., 2000. The economic value of controlling an invasive shrub. Ambio 29, 462–467.
9
Contents lists available at ScienceDirect
Int J Appl Earth Obs Geoinformation journal homepage: www.elsevier.com/locate/jag
Rapid assessment of juniper distribution in prairie landscapes of the northern Great Plains
T
Kyle D. Kaskiea, Michael C. Wimberlyb, , Peter J. Baumana ⁎
a b
Department of Natural Resource Management, South Dakota State University, Brookings, SD, 57007, USA Department of Geography and Environmental Sustainability, University of Oklahoma, Norman, OK, 73019, USA
ARTICLE INFO
ABSTRACT
Keywords: Spectral mixture analysis Matched filtering Landsat Accuracy assessment Tree invasion Eastern redcedar
Woody plant species including eastern redcedar (Juniperus virginiana) and Rocky Mountain juniper (Juniperus scopulorum) are expanding throughout the prairie ecosystems of the Great Plains because of fire suppression, land management practices, and climate change. Juniper encroachment threatens native grasslands by altering soil characteristics, limiting herbaceous biomass, hindering native plant regeneration, and reducing rangeland productivity. Existing land cover products do not effectively characterize the distribution of juniper over a range of densities, making it difficult to assess the scale of the problem. We evaluated a method for rapid classification and mapping of juniper using matched filtering with Landsat 8 snow-covered and snow-free winter imagery (January-March), and snow-free spring imagery (April–June) for 2015–2016. We used data from two path/rows (29/30 and 30/30) in southeastern South Dakota and northeastern Nebraska (approximately 23,000 km2). In both path/rows, we found that snow-covered winter images increased contrast between juniper and the image background and resulted in the highest overall classification accuracies of 94.5% and 88.9% for juniper densities above 15%, compared to 91.4% and 85.7% for snow-free winter imagery and 57.8% and 74.1% for growing season imagery. Using winter imagery, we successfully captured pixels containing juniper density above 50% with ≥90% detection probability. However, the true positive rate dropped to less than 50% once juniper density fell below 20%. We identified 84 791 ha within the study area occupied by juniper (3.6% of the total area), including 27 504 ha in deciduous forests (33% of deciduous forest area) and 38 738 ha in grasslands (6% of grassland area).
1. Introduction The Great Plains has been recognized as one of North America’s most endangered ecosystems (Samson and Knopf, 1996). A significant factor driving loss of native ecosystems has been the expansion of agriculture and the resulting conversion of grassland to cultivated cropland (Samson et al., 2004; Wright and Wimberly, 2013; Wimberly et al., 2017). However, the encroachment of woody plants, particularly eastern redcedar (Juniperus virginiana) and Rocky Mountain juniper (Juniperus scopulorum), is also contributing to the widespread degradation and loss of prairie ecosystems in the Great Plains. Historically, the geographic extent of these species (collectively referred to as “juniper” hereafter) was greatly restricted by frequent fires, but in the modern landscape juniper has expanded its range by encroaching into grasslands (Briggs et al., 2002a; Twidwell et al., 2013). A long history of fire suppression and other land use practices have contributed to increases in juniper cover throughout the Great Plains (Briggs et al.,
⁎
2002a). In addition, environmental fluctuations, particularly droughts, can give junipers a competitive advantage over other species and facilitate their expansion (Volder et al., 2013). Juniper encroachment has already overwhelmed much of the southern Great Plains (Norris et al., 2001; Starks et al., 2014) and is spreading rapidly in the North (Meneguzzo and Liknes, 2015; Pierce and Reich, 2010). Woody plant encroachment alters native grasslands by changing soil characteristics, limiting herbaceous biomass, hindering regeneration of native grassland species, and reducing rangeland forage production (Briggs et al., 2002b; Gehring and Bragg, 1992; McKinley and Blair, 2008). Juniper encroachment also results in degradation of habitat quality for grassland birds and other native wildlife species (Coppedge et al., 2004). Management practices such as mechanical removal, herbicide application, and prescribed burning are commonly used to reduce tree cover and restore grassland communities (Wilson and Schmidt, 1990). However, these management activities can be time consuming and costly, and their effectiveness generally decreases as
Corresponding author. E-mail address: [email protected] (M.C. Wimberly).
https://doi.org/10.1016/j.jag.2019.101946 Received 17 June 2019; Received in revised form 13 August 2019; Accepted 16 August 2019 Available online 26 August 2019 0303-2434/ © 2019 Elsevier B.V. All rights reserved.
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
tree size and stand density increase (Bidwell et al., 2002; Buehring et al., 1971; Ortmann et al., 1998). Detailed and accurate maps of juniper distribution can aid in determining the extent of juniper cover, assessing the ecological impacts of juniper encroachment, and developing strategies for prioritizing juniper management across large regions encompassing thousands to millions of hectares. There are multiple sources of data on juniper extent and dynamics, each with particular advantages and limitations. The Forest Inventory and Analysis (FIA) program of the USDA Forest Service collects field data on a sample of one-acre forestland plots within a state, allowing computation of yearly forest type estimates and providing a complete statewide inventory at five year intervals. Meneguzzo and Liknes (2015) used these data to show that over a seven-year time span, (2005–2012) the juniper forest type increased by 116 000 ha in the central United States, resulting in an annual loss of 16 500 ha of nonforestland to juniper encroachment between 2007 and 2012. FIA data have also been combined with Moderate Resolution Imaging Spectroradiometer (MODIS) 250 m imagery to interpolate and produce live volume and density maps of juniper (Meneguzzo et al., 2008; Simonsen et al., 2015). Although these coarse-grained maps and area estimates are suitable for general assessments of juniper distribution over large areas, they do not precisely depict all juniper locations in the landscape. Accurate higher-resolution maps are needed to make large-area assessments and inform site-specific management decisions. Previous studies have used a variety of remote sensing data, including very high spatial resolution (VHSR) aerial imagery (Anderson and Cobb, 2004; Poznanovic et al., 2014), hyperspectral imagery (Wylie et al., 2000), and fusion of Landsat Thematic Mapper (TM) imagery with Light Detection and Ranging (LiDAR) data (Sankey et al., 2010). A more recent study has shown the potential for mapping eastern redcedar at a large scale with the combination of L-band Synthetic Aperture Radar (SAR) data and Landsat Thematic Mapper/Enhanced Thematic Mapper Plus (TM/ETM +) imagery (Wang et al., 2017). Although these studies have provided useful results, many of the data sources have limited spatial or temporal coverage, making these approaches difficult to apply to other study areas. Landsat imagery is a freely accessible data source that allows for large-scale mapping and continuous monitoring of woody plant encroachment (Sankey and Germino, 2008; Venter et al., 2018; Wang et al., 2017; Yang et al., RSE 2012). Since 1984, the Landsat program has provided 30-m multispectral images at a 16-day temporal resolution (Wulder et al., 2006). The Landsat 8 mission, launched in 2013, features a global acquisition strategy, has improved radiometric resolution compared to previous Landsat sensors, and eliminates the data gaps associated with the scan-line corrector failure on Landsat 7. One limitation of Landsat is that the pixel size is larger than most individual trees, so that pixels often contain a mixture of trees and herbaceous grassland vegetation. For this reason, sub-pixel methods such as spectral mixture analysis have been utilized for classifying juniper encroachment with Landsat imagery (Sankey and Germino, 2008). To generate juniper maps that are useful for management purposes, there is a need for replicable classification methods that are straightforward to implement and use readily available open source data to support high-accuracy mapping. To address this need, we investigated the classification of juniper in a prairie landscape using a linear spectral unmixing method with Landsat 8 satellite imagery. Our objectives were to: 1) evaluate the potential for juniper detection using partial unmixing techniques with Landsat 8 OLI (Operational Land Imager) imagery; 2) compare classification accuracies for imagery from different seasons with distinctive spectral characteristics (growing season, non-growing season with consistent snow coverage, and non-growing season imagery with no snow coverage); 3) assess classification accuracy for varying levels of juniper density; and 4) compare patterns of juniper distributions with generic land cover classes from the National Land Cover Database to determine the new information gained from the juniper map.
2. Material and Methods 2.1. Study Area The study area covered 14 contiguous counties bordering the Missouri River, including nine counties in southeastern South Dakota and five counties in northeastern Nebraska (Fig. 1). This area has a Köppen climate classification of humid continental (Dfa) (Kottek et al., 2006), with an annual mean temperature varying between 6–11 °C and an annual precipitation varying between 498–796 mm from 1981 through 2010. Steeply sloped drainages disrupt a flat to rolling topography with land cover comprised mostly of cropland (48%) and herbaceous grasslands (39%). The primary land uses include the agricultural production of corn, soybeans, and wheat as well as cattle grazing. Common native grassland vegetation consists of mixed-grass prairie species such as little bluestem (Schizachyrium scoparium), big bluestem (Andropogon gerardii), western wheatgrass (Pascopyrum smithii), sideoats grama (Bouteloua curtipendula), and green needlegrass (Nassella viridula). Woodlands are primarily found near drainages and riparian lowlands with the exception of small groves scattered across the prairie uplands. Common deciduous tree species include the plains cottonwood (Populus deltoids), bur oak (Quercus macrocarpa), green ash (Fraxinus pennsylvanica) and American elm (Ulmus americana). Juniper species such as Rocky mountain juniper (Juniperus scopulorum) and eastern redcedar (Juniperus virginianai) are also common (Barker and Whitman, 1988). 2.2. Data sources We obtained Landsat 8 Operational Land Imager (OLI) level-2 surface reflectance data for path/row 29/30 and 30/30 (Fig. 1) containing minimal cloud cover (< 10%) through the U.S. Geological Survey (USGS) EarthExplorer online tool (USGS Earth Explorer, 2017). These data were generated using the Landsat Surface Reflectance Code (LaSRC). We inspected each image and its associated pixel quality band and selected three images with minimal cloud cover that covered the study area for each path/row. We chose images from different seasons with distinctive spectral characteristics, including one image from the growing season (April through July) with mostly green vegetation, and two images from the non-growing season (January through March). For the non-growing season, we selected one image with uniform snow coverage and one image with no snow coverage that was predominantly senesced vegetation and soil. Snow cover has been shown to reduce spectral variability among ground surface targets when using Landsat sensor data while also increasing contrast between white snow and forest elements foliage, bark, and shadows above the snow cover (Wolter et al., 2008, 2012). Table 1 shows the final six images chosen for analyses. Once we identified suitable image dates, we extracted surface reflectance data for six bands: band 2 (0.435–0.451 μm, blue), band 3 (0.533–0.590 μm, green), band 4 (0.636–0.673 μm, red), band 5 (0.851–.0879 μm, near infrared), band 6 (1.566–1.651 μm, shortwave infrared), and band 7 (2.107–2.294 μm, shortwave infrared). To reduce spectral misclassification, we constructed a mask for each image that excluded any permanent water body features (i.e. streams, rivers, ponds and lakes). We identified water bodies using the USGS National Hydrography Dataset (NHD), obtained through Geospatial Data Gateway (NRCS, 2017). 2.3. Juniper classification We performed juniper classification using a matched filtering approach with ENVI version 5.4 (Exelis Visual Information Solutions, Boulder, Colorado). Matched filtering is a partial unmixing procedure that incorporates user-defined endmembers to maximize the response 2
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 1. Study area composed of nine counties in southeastern South Dakota and five counties in northeastern Nebraska. Landsat 8 path/rows: 30/30 (a) and 29/30 (b) cover the 14 contiguous counties. Points indicate the locations of visually interpreted juniper presence/absence pixels for validation along with randomly generated and opportunistically chosen field investigation sites.
of known spectral signatures while suppressing unknown background signatures. The required inputs for matched filtering included the spectral signature of a pure endmember, the selection of bands to use with the algorithm, and the selection of a threshold to classify pixels as juniper on non-juniper. Matched filtering requires only the input of the desired endmember or cover type that the user wishes to identify. This approach differs from conventional linear spectral mixture analysis, which requires an input of all known endmembers. Previous partial unmixing research has indicated that the mean of manually selected endmembers containing a high percentage of target cover outperformed extreme or variant n-dimensional visualizer (ND-V) endmember pixels and the mean of all NDV endmember pixels (Sankey and Glenn, 2011). Therefore, we selected ten pixels that were predominantly juniper within each path/row (29/ 30 and 30/30). This process allowed the average spectral signature to be obtained, which we used as the endmember input for the matched filtering analyses.
Previous studies and preliminary observations suggested that a combination of bands 2–5 and band 7 allowed for the best spectral separation between juniper and background materials (Vikhamar and Solberg, 2003). We therefore used these bands to generate a matched filtering image containing values that represented the relative degree of match for each pixel. A value of one signified a perfect match while values closer to zero indicated background or non-target materials. Once we completed the matched filtering analysis, we developed a binary juniper classification map by designating a threshold for the matched filtering values to discriminate juniper from non-juniper pixels. We determined a threshold for each image by sampling a group of user-defined pixels that contained a high percentage (> 50%) of juniper cover. We then computed the mean matched filtering value and standard deviation for the sampled pixels and assigned two negative standard deviations from the mean as the threshold. This threshold was applied to produce a final classified map that contained pixels representing juniper and non-juniper.
Table 1 Landsat 8 OLI Surface Reflectance Level-2 imagery used for the classification of juniper. Product Identifier
Date
Path/row
Cloud cover (%)
Snow cover (%)
LC08_L1TP_029030_20150107_20170302_01_T2 LC08_L1TP_029030_20150328_20170227_01_T1 LC08_L1TP_029030_20160602_20170223_01_T1 LC08_L1TP_030030_20160101_20180131_01_T1 LC08_L1TP_030030_20160305_20170224_01_T1 LC08_L1TP_030030_20160524_20180131_01_T1
2015/01/07 2015/03/28 2016/06/02 2016/01/01 2016/03/05 2016/05/24
29/30 29/30 29/30 30/30 30/30 30/30
1.55 1.11 7.41 0.05 0.45 0.00
94.75 0.00 0.00 97.91 0.00 0.00
3
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 2. Qualitative assessment of juniper classification maps produced by Landsat 8 imagery during (a) the non-growing season with no snow coverage, (b) the nongrowing season with consistent snow coverage, and (c) the growing season (c). Pixels classified as juniper are displayed over 2016 NAIP imagery.
2.4. Accuracy assessment
characteristics (e.g. land use, species composition). We also took additional photos at opportunistic stops. We collected data for 252 sites and obtained 1271 photos distributed throughout our study area (Fig. 1).
To assess accuracy, we used a random stratified sampling design that allowed us to determine the classification accuracy over a range of juniper densities. We specified four strata, which included closed canopy woodlands, buffered closed canopy woodlands, planted shelterbelts, and other non-woodland areas. We digitized the closed canopy woodlands and planted shelterbelts in ArcGIS 10.5 (ESRI Inc., Redlands CA) following the guidelines presented in Bauman et al. (2016) using very high spatial resolution (VHSR) 60 cm National Agricultural Imagery Program (NAIP), 2014 and 2016 aerial imagery. In addition to obtaining samples of dense woodland cover, we also sampled open canopied woodlands and sparse trees. To do so, we placed a 90-m buffer around the digitized closed canopy stratum using ArcGIS. Based on visual observations, we determined that this distance encompassed many areas with low densities of regenerating juniper. Once we defined all the strata within the study area, we generated random points using ArcGIS. We referenced each random point to a Landsat pixel by converting it to a 30 × 30 m polygon and snapping it to the Landsat 8 grid. Then, we characterized each polygon using a combination of VHSR imagery, which included NAIP 2016 and other sources of winter imagery accessed through Google Earth from 2013 to 2017. Data collected for each sample included the land cover type, juniper presence/absence, and the percent of juniper cover within the 30 × 30 m polygon. We delineated a total of 1,643 juniper presence and 2,273 juniper absence points for accuracy assessment covering a range of juniper densities (Fig. 1). We conducted field investigations during October 2017 to obtain ground reference data and gain a better understanding of juniper distribution over the study area. Random locations were selected along road corridors and used to assess 0.5 km driven transects. Each transect was divided into three stops, and each stop was divided into a left and right side (six stops per transect). At each stop, we captured a photo with a GPS enabled digital camera and recorded vegetation and land cover
2.5. Comparison with land cover The National Land Cover Database (NLCD) provides a consistent, national-level map of generic land cover classes at a 30-m spatial resolution that is updated on a regular basis (Homer et al., 2015). To determine whether the juniper map developed in this study provided novel information not already available through the NLCD, and to provide a broader context for interpreting the juniper patterns, we conducted a cross-tabulation of juniper presence/absence with the 2016 NLCD. We assessed the degree to which the juniper maps corresponded with evergreen and mixed forest mapped in the NLCD, and the distribution of juniper across the various NLCD classes. 3. Results We produced a total of six Landsat 8 juniper classification maps under three image conditions (consistent snow coverage during the non-growing season, no snow coverage during the non-growing season, and growing season) for two path/rows: 29/30 and 30/30. A visual assessment using VHSR imagery and field photographs found a reasonably accurate representation of juniper for all images in the nongrowing season (Fig. 2a, b). However, misclassification of non-juniper (i.e. cultivated fields and wetlands) as juniper was observed more frequently in images containing no snow coverage (Fig. 2a) than it was in images containing consistent snow coverage (Fig. 3b). Maps based on growing season imagery appeared to have a higher misclassification of both juniper and non-juniper sites (Fig. 2c), which was confirmed by the quantitative accuracy assessments. 4
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 3. Pixel-level accuracy assessment for the classification of juniper by density with non-growing season imagery. Path/row: (a) 29/30 and (b) 30/30.
3.1. Accuracy assessment
values of 92.1% and 84.7% for path/rows 29/30 and 30/30 compared to 83.2% and 76.3% for snow-covered non-growing season imagery and 70.9% and 70.0% for non-growing season imagery without snow cover. However, both growing season images had much lower user accuracies (the probability that a pixel mapped as juniper actually represents juniper in the reference dataset) of 38.7% and 67.4% compared to 96.3% and 99.2% for snow-covered non-growing season and 96.8% and 98.3% for non-growing season imagery without snow cover. Because of the low user accuracies for the maps based on growingseason imagery, we focused the next phase of the analysis on the maps
The classified maps based on imagery with consistent snow coverage during the non-growing season had the highest accuracy statistics for both path/row combinations (Table 2). These maps had overall accuracy (OA) of 94.5% and 88.9% for path/rows 29/30 and 30/30 compared to 91.4% and 85.7% for non-snow coverage imagery during the non-growing season and 57.8% and 74.1% for the growing season. Producer accuracies (the probability that a reference pixel of juniper is correctly classified) were highest for the growing season imagery with
Table 2 Accuracy assessment for six Landsat 8 scenes based on classified juniper presence and absence pixels. Validation points were considered to have juniper present if they had a juniper cover greater than 15%. Shown underlined, overall accuracy (OA) represents the total classification accuracy for both juniper presence and juniper absence pixels. Path/Row
Condition
Classified pixel data
Reference pixel data
User Accuracy (UA)
Juniper absence
Juniper presence
Row totals
29/30
Snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
887 11 898 0.9878
57 283 340 0.8324
944 294 1238
0.9396 0.9626
29/30
Non-snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
890 8 898 0.9911
99 241 340 0.7088
989 249 1238
0.8999 0.9679
29/30
Growing
Juniper absence Juniper presence Column totals Producer accuracy (PA)
403 495 898 0.4488
27 313 340 0.9206
430 808 1238
30/30
Snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
978 5 983 0.9949
198 639 837 0.7634
1176 644 1820
30/30
Non-snow
Juniper absence Juniper presence Column totals Producer accuracy (PA)
973 10 983 0.9898
251 586 837 0.7001
1224 596 1820
Juniper absence Juniper presence Column totals Producer accuracy (PA)
640 343 983 0.6511
128 709 837 0.8471
768 1052 1820
30/30
Growing
5
0.9101
0.8500 0.9372 0.3874 0.6847 0.8316 0.9922 0.8792 0.7949 0.9832 0.8450 0.8333 0.6740 0.7491
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 4. Examples of juniper density at (a) 1–10, (b) 10–20, (c) 20–30, (d) 30–40, (e) 40–50, (f) 50–60, (g) 60–70, (h) 70–80, (i) 80–90, and (j) 90–100%. 30 × 30 m polygons are overlaid on NAIP and Google Earth imagery.
based on non-growing season imagery. The true positive rate (calculated as the percent of juniper validation points that were correctly classified) of matched filtering classification of juniper for non-growing season imagery was assessed at ten juniper density levels (Figs. 3 and 4). The true positive rate of juniper for both image conditions in the non-growing season was 90% or higher (Fig. 3) for pixels containing juniper density above 50% (Fig. 4f–j). When juniper density was below 50%, the true positive rate varied more between each path/row and image condition (snow versus no snow cover) than for pixels containing higher densities of juniper. Imagery with consistent snow coverage produced higher true positive rates at lower levels of juniper coverage than images with no snow coverage (Fig. 3). The true positive rate dropped below 50% for non-snow covered images when juniper densities were at 20–30% (Fig. 4c; path/row: 29/30 and 30/30; 47% and 40%) and for snow-covered images at 10–20% (Fig. 4b; 48% and 37%). For pixels with the lowest level of juniper cover ranging from 1 to 10% (Fig. 4a), true positive rates were extremely low, with a probability of detection less than 13% for all images.
4. Discussion 4.1. Map accuracy Our visual and quantitative accuracy assessments demonstrate that matched filtering of Landsat 8 OLI imagery with consistent snow coverage during the non-growing season is an effective technique for mapping landscape-level patterns of juniper across large areas in the northern Great Plains. Practical advantages are that this technique requires spectral data from a relatively small number of juniper-dominated pixels to characterize the juniper endmember and can be implemented with a single Landsat scene. If a suitable cloud-free, snowcovered imagery is available for the location and time of interest, it is possible to rapidly and efficiently generate juniper distribution maps. Using the matched filtering approach, we were able to obtain accuracies comparable to other juniper-mapping studies that used multiple data sources (Sankey et al., 2010; Wang et al., 2017) or hyperspectral data (Wylie et al., 2000). When images from the non-growing season were used, the resulting maps captured pixels containing juniper density above 50% with a 90% and greater detection probability, and we also obtained a true positive rate of 50% or greater for pixels containing juniper density above 15%. Wang et al. (2017) similarly achieved a 90% detection probability for pixels containing eastern redcedar (Juniperus virginiana) density above 60% and a 30% detection probability when pixel densities where between 10 to 20% using a combination of multi-temporal Landsat imagery and L-band SAR imagery. We were able to obtain comparable classification accuracies for a single-date map based on a single Landsat 8 image. The Landsat-based juniper classification maps produced higher overall classification accuracies for images during the non-growing season compared to images obtained during the growing season. This difference can be attributed to smaller differences in spectral signatures between juniper and other types of vegetation and woodlands during the growing season compared to the winter months, when juniper was the only green vegetation (Wang et al., 2017). We were also able to obtain a better spectral separation of juniper and non-juniper when images containing consistent snow coverage were used compared to images containing no snow coverage during the non- growing season. The snow-covered pixels contained higher reflectance values in the visible wavelengths and near-infrared wavelengths (Dozier and Painter, 2004; Vikhamar and Solberg, 2003). This bright background provided a strong contrast to juniper, which had very low reflectance in the visible
3.2. Juniper patterns and land cover associations A final juniper map was created by mosaicking the juniper classification maps produced from snow-covered imagery for path/rows 29/30 and 30/30. This map, shown in Fig. 5a, identified areas of planted juniper while also detecting areas of naturally-regenerated juniper (Fig. 5b, c). Examination of the final map also showed that the matched filtering technique with snow-covered imagery could effectively distinguish juniper cover from surrounding deciduous woodlands (Fig. 5d, e). The juniper map identified 84 791 ha occupied by juniper (3.6% of the total area). Most of the evergreen forest (89%) and mixed forest (61%) classes from the NLCD were occupied by juniper in our map (Table 3). In addition, 33% of deciduous forest, 17% of woody wetlands, 10% of shrublands, and 6% of grasslands were occupied by juniper. However, only 11% of the juniper area was classified as either evergreen or mixed forest in the NLCD. In addition, 32% of the juniper area (27 504 ha) was classified as deciduous forest and 46% of the juniper area (38 738 ha) was classified as grassland. Only small percentages of the juniper area (< 3%) were classified as other NLCD classes. Overall, the NLCD evergreen and mixed forest classes substantially underrepresented the extent of juniper, which is the most widespread evergreen tree species within the study area. 6
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
Fig. 5. (a) Final juniper classification map. (b, c) Close-up maps from region 2 showing juniper spread from planted shelterbelts. (d,e) Close-up maps from region 1 showing the discrimination of juniper from deciduous trees. A Google Earth high-resolution image dated 4/11/2016 shows junipers as green while leaf-off deciduous trees are gray-brown. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
wavelengths and near-infrared wavelengths compared to the snowcovered pixels. The snow cover aided in the classification of juniper by reducing variation in the background non-juniper matrix, permitting our material of interest (i.e., juniper) to be the primary driver of the scene covariance (Boardman et al., 1995). However, a limitation to the use of snow-covered images is that they may not be available for every year and location of interest, and they will become progressively more difficult to obtain in warmer locations where snowfall is less common. The pixel-level accuracy assessments showed a strong trend associated with juniper density, where probability of correctly classifying a pixel as containing juniper increased with juniper density. This
observation was also noted in the recent study of Wang et al. (2017), who suspected that incorrectly classified low-density juniper pixels were influenced by juniper height and the omission of woodlands that did not meet their definition of a forest. In our study the loss of accuracy for low-density juniper, particularly at levels of cover less than 20%, can be attributed to the weaker spectral signatures of pixels containing only a small fraction of juniper cover. Low-density juniper pixels contain more non-juniper vegetation in a heterogeneous background, whereas pixels with higher juniper density have less background vegetation (Briggs et al., 2002b; Gehring and Bragg, 1992). These differences influence the partial spectral unmixing classification because
Table 3 Cross-tabulated areas from the overlay of the classified juniper map with land cover classes the 2016 National Land Cover Database (NLCD). NLCD Class
Non-Juniper (ha)
Juniper (ha)
Total (ha)
% NLCD Classified as Juniper
% Juniper in NLCD Class
Water Developed (Open) Developed (Low) Developed (Medium) Developed (High) Barren Deciduous Forest Conifer Forest Mixed Forest Shrub Grassland Pasture/Hay Crop Woody Wetland Emergent Wetland Total
77 174 65 508 12 221 3 259 935 1 869 54 611 995 489 5 315 645 345 201 900 1 169 616 10 773 38 830 2 288 840
365 965 297 37 4 19 27 504 8 211 773 606 38 738 1 818 2 357 2 216 880 84 791
77 540 66 473 12 518 3 296 939 1 888 82 115 9 206 1 262 5 920 684 083 203 718 1 171 973 12 988 39 710 2 373 630
0.47 1.45 2.37 1.14 0.43 1.01 33.49 89.19 61.28 10.23 5.66 0.89 0.20 17.06 2.22
0.43 1.14 0.35 0.04 0.00 0.02 32.44 9.68 0.91 0.71 45.69 2.14 2.78 2.61 1.04
7
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al.
non-juniper spectral signatures mask the spectral signatures of junipers within the pixel. If the classification threshold was decreased, the probability of positively identifying lower density pixels would increase, but at the expense of increasing the classification of more nonjuniper sites as juniper (i.e., false positives). The low classification accuracy for low-density juniper pixels is a limitation of our mapping approach. Classifying these low-density juniper pixels can be particularly important because these are the areas where juniper removal via fire or mechanical cutting is most cost-effective and where grassland vegetation is still present and there is the greatest potential for prairie restoration. Because of the constraints imposed by Landsat’s 30-m resolution, additional data sources and methods will be needed to reliably identify low-density juniper. One possibility is to use very high-resolution aerial or satellite imagery to identify and map individual juniper trees (Falkowski et al., 2017; Ozdemir, 2008; Poznanovic et al., 2014). Another option could be to predict areas that are likely to support low densities of juniper regeneration based on the locations of juniper seed sources and environmental settings conducive to juniper establishment (Greene and Knox, 2014; Weisberg et al., 2007).
communities (Knapp et al., 2008; Ratajczak et al., 2012; Van Auken, 2009). Accurate and timely data on the distribution of woody plant encroachment is essential for assessing the economic and ecological impacts of woody encroachment (Anadón et al., 2014; Zavaleta, 2000) and will assist in establishing and implementing the appropriate management measures (Bidwell et al., 2002; Ortmann et al., 1998; Smith, 2011). Juniper species are the main woody plant species threatening grassland habitats in our region (Engle et al., 2008), and we were able to develop maps that accurately depict the distribution of juniper by using partial unmixing techniques with Landsat 8 OLI imagery. This workflow provides a straightforward means for mapping juniper distribution over a large scale, allowing managers to quantify the regional impact of juniper expansion and identify sites where further encroachment is likely in the future. Acknowledgements We thank Izaya Numata and Justin Davis for advice on remote sensing and analytical methods. Joshua Leffler provided helpful comments on an earlier draft of the manuscript. This research was supported by the Geospatial Sciences Center of Excellence at South Dakota State University.
4.2. Applications of the juniper distribution map The methods developed in this study allows for the mapping of juniper distribution across large regions and can facilitate regional reassessment of management polices (Roberts et al., 2018). Juniper maps can also aid in monitoring to determine where juniper has encroached into environmentally sensitive areas, supporting management decisions that can help prevent of future grassland loss (Bauman et al., 2016; Wimberly et al., 2018). The juniper maps developed through this research will also establish a baseline for future studies. Although this study focused on only two Landsat path/rows and a single date, the same partial unmixing techniques could be applied across larger extents and over multiple time periods. The use of remote sensing to address the dynamic encroachment of juniper at a large scale has only recently been investigated in the southern Great Plains (Wang et al., 2018), but is yet to be explored in the northern Great Plains. With a better understanding of the drivers associated with juniper encroachment, current ecological knowledge (Briggs et al., 2002b; Pierce and Reich, 2010) and the development of accurate juniper maps, we can manage encroaching juniper species more efficiently by targeting proactive management strategies and restoration activities (Greene and Knox, 2014). The juniper map yielded new information that complements the data provided by the NLCD, a standard national-level land cover data product. The NLCD has been developed to provide accurate mapping of generic land cover classes at regional to national levels (Homer et al., 2015) but is not optimal for capturing sub-pixel land-cover features such as juniper encroachment. Although the NLCD forest classes are specified to identify areas with more than 20% trees, the total area of evergreen and mixed forests was much smaller than the area occupied by juniper in our study area. Most of the juniper areas were located in grasslands with partial juniper cover or in deciduous forest stands with a juniper component. Our juniper map provides unique regionally-relevant information that complements the NLCD, and we have demonstrated that combining these two data sources can provide valuable information about the extent and pattern of juniper encroachment. For example, we estimated that 6% of the grasslands in the study area, encompassing 38 783 ha, are currently being impacted by juniper encroachment, and the juniper map can be used to find the locations of these areas. In contrast, interpreting the NLCD conifer and mixed forest classes as a measurement of juniper distribution would greatly underestimate the impacted area. Although this study focused primarily on mapping juniper encroachment in a portion of the northern Great Plains, woody plant encroachment is also threatening other North American grassland
References Anadón, J.D., Sala, O.E., Turner, B.L., Bennett, E.M., 2014. Effect of woody-plant encroachment on livestock production in North and South America. Proc. Natl. Acad. Sci. U.S.A. 111, 12948–12953. Anderson, J.J., Cobb, N.S., 2004. Tree cover discrimination in panchromatic aerial imagery of pinyon-juniper woodlands. Photogramm. Eng. Remote Sensing 70, 1063–1068. Barker, W.T., Whitman, W.C., 1988. Vegetation of the northern Great Plains. Rangelands 10, 266–272. Bauman, P., Carlson, B., Butler, T., 2016. Quantifying Undisturbed (Native) Lands in Eastern South Dakota: 2013. South Dakota State University Extension, Brookings, SD. Bidwell, T.G., Weir, J.R., Engle, D.M., 2002. Eastern Redcedar Control and Management: Best Management Practices to Restore Oklahoma’s Ecosystems. Oklahoma Cooperative Extension Service Fact Sheet NREM-2876. Briggs, J.M., Knapp, A.K., Brock, B.L., 2002a. Expansion of woody plants in tallgrass prairie: a fifteen-year study of fire and fire-grazing interactions. Am. Midl. Nat. 147, 287–294. Briggs, J.M., Hoch, G.A., Johnson, L.C., 2002b. Assessing the rate, mechanisms, and consequences of the conversion of tallgrass prairie to Juniperus virginiana forest. Ecosystems 5, 578–586. Boardman, J.W., Kruse, F.A., Green, R.O., 1995. Mapping target signatures via partial unmixing of AVIRIS data. In: Proceedings of the Fifth JPL Airborne Earth Science Workshop. Pasadena, CA. pp. 95–101. Buehring, N., Santelmann, P.W., Elwell, H.M., 1971. Responses of eastern red cedar to control procedures. J. Range Manag. 24, 378–382. Coppedge, B.R., Engle, D.M., Masters, R.E., Gregory, M.S., 2004. Predicting juniper encroachment and CRP effects on avian community dynamics in southern mixed-grass prairie. USA. Biol. Conserv 115, 431–441. Dozier, J., Painter, T.H., 2004. Multispectral and hyperspectral remote sensing of alpine snow properties. Annu. Rev. Earth Planet. Sci. 32, 465–494. Engle, D.M., Coppedge, B.R., Fuhlendorf, S.D., 2008. From the dust bowl to the green glacier: human activity and environmental change in Great plains grasslands. In: van Auken, O. (Ed.), Western North American Juniperus Communities. Springer, New York, pp. 253–271. Falkowski, M.J., Evens, J.S., Naugle, D.E., Hagen, C.A., Carleton, S.A., Maestas, J.D., Khalyani, A.H., Poznanovic, A.J., Lawrence, A.J., 2017. Mapping tree canopy cover in support of proactive prairie grouse conservation in western North America. Rangel. Ecol. Manag. 70, 15–24. Gehring, J.L., Bragg, T.B., 1992. Changes in prairie vegetation under eastern red cedar (Juniperus virginiana L.) in an eastern Nebraska bluestem prairie. Am. Midl. Nat. 128, 209–217. Greene, S.L., Knox, J.C., 2014. Coupling legacy geomorphic surface facies to riparian vegetation: assessing red cedar invasion along the Missouri River downstream of Gavins Point dam, South Dakota. Geomorphology 204, 277–286. Homer, C., Dewitz, J., Yang, L., Jin, S., Danielson, P., Xian, G., Coulston, J., Herold, N., Wickham, J., Megown, K., 2015. Completion of the 2011 National Land Cover Database for the conterminous United States–representing a decade of land cover change information. Photogramm. Eng. Remote Sensing 81, 345–354. Knapp, A.K., Briggs, J.M., Collins, S.L., Archer, S.R., Bret-Harte, M.S., Ewers, B.E., Peters, D.P., Young, D.R., Shaver, G.R., Pendall, E., Cleary, M.B., 2008. Shrub encroachment in North American grasslands: shifts in growth form dominance rapidly alters control of ecosystem carbon inputs. Glob. Chang. Biol. 14, 615–623. Kottek, M., Grieser, J., Beck, C., Rudolf, B., Rubel, F., 2006. World map of the KöppenGeiger climate classification updated. Meteorol. Z. 15, 259–263.
8
Int J Appl Earth Obs Geoinformation 83 (2019) 101946
K.D. Kaskie, et al. McKinley, D.C., Blair, J.M., 2008. Woody plant encroachment by Juniperus virginiana in a mesic native grassland promotes rapid carbon and nitrogen accrual. Ecosystems 11, 454–468. Meneguzzo, D.M., Butler, B.J., Crocker, S.J., Haugen, D.E., Moser, W.K., Perry, C.H., Wilson, B.T., Woodall, C.W., 2008. Nebraska’s forests, 2005. Resour. Bull. NRS-27. U.S. Department of Agriculture Forest Service, Northern Research Station, Newtown Square, PA. Meneguzzo, D.M., Liknes, G.C., 2015. Status and trends of eastern redcedar (Juniperus virginiana) in the central United States: analyses and observations based on Forest Inventory and Analysis data. J. For. 113, 325–334. Norris, M.D., Blair, J.M., Johnson, L.C., 2001. Land cover change in eastern Kansas: litter dynamics of closed-canopy eastern redcedar forests in tallgrass prairie. Can. J. Bot. 79, 214–222. Ortmann, J., Stubbendieck, J., Masters, R.A., Pfeiffer, G.H., Bragg, T.B., 1998. Efficacy and costs of controlling eastern redcedar. J. Range Manag. 51, 158–163. Ozdemir, I., 2008. Estimating stem volume by tree crown area and tree shadow area extracted from pan-sharpened Quickbird imagery in open Crimean juniper forests. Int. J. Remote Sens. 29, 5643–5655. Pierce, A.M., Reich, P.B., 2010. The effects of eastern red cedar (Juniperus virginiana) invasion and removal on a dry bluff prairie ecosystem. Biol. Invasions 12, 241–252. Poznanovic, A.J., Falkowski, M.J., Maclean, A.L., Smith, A., Evans, J.S., 2014. An accuracy assessment of tree detection algorithms in juniper woodlands. Photogramm. Eng. Remote Sensing 80, 627–637. Ratajczak, Z., Nippert, J.B., Collins, S.L., 2012. Woody encroachment decreases diversity across North American grasslands and savannas. Ecology 93, 697–703. Roberts, C.P., Uden, D.R., Allen, C.R., Twidwell, D., 2018. Doublethink and scale mismatch polarize policies for an invasive tree. PLoS One 13, e0189733. Samson, F.B., Knopf, F.L. (Eds.), 1996. Prairie Conservation: Preserving North America’S Most Endangered Ecosystem. Island Press, Washington D.C. Samson, F., Knopf, F., Ostlie, W., 2004. Great Plains ecosystems: past, present, and future. Wildl. Soc. Bull. 32, 6–15. Sankey, T.T., Germino, M.J., 2008. Assessment of juniper encroachment with the use of satellite imagery and geospatial data. Rangel. Ecol. Manag. 61, 412–418. Sankey, T.T., Glenn, N., Ehinger, S., Boehm, A., Hardegree, S., 2010. Characterizing western juniper expansion via a fusion of Landsat 5 Thematic Mapper and lidar data. Rangel. Ecol. Manag. 63, 514–523. Sankey, T., Glenn, N., 2011. Landsat-5 TM and lidar fusion for sub-pixel juniper tree cover estimates in a western rangeland. Photogramm. Eng. Remote Sensing 77, 1241–1248. Simonsen, V.L., Fleishmann, J.E., Whisenhunt, D., Volesky, J.D., Twidwell, D., 2015. Act Now or Pay Later: Evaluating the Cost of Reactive Versus Proactive Eastern Redcedar Management. University of Nebraska-Lincoln Extension Report EC1784. Smith, S., 2011. Eastern Red-cedar: Positives, Negatives and Management. Samuel Roberts Noble Foundation, Ardmore, OK. Starks, P.J., Venuto, B.C., Dugas, W.A., Kiniry, J., 2014. Measurements of canopy interception and transpiration of eastern redcedar grown in open environments. Environ. Nat. Resour. Res. 4, 103–122. Twidwell, D., Rogers, W.E., Fuhlendorf, S.D., Wonkka, C.L., Engle, D.M., Weir, J.R., Kreuter, U.P., Taylor, C.A., 2013. The rising Great Plains fire campaign: citizens’
response to woody plant encroachment. Front. Ecol. Environ. 11, e65–e71. Van Auken, O.W., 2009. Causes and consequences of woody plant encroachment into western North American grasslands. J. Environ. Manage. 90, 2931–2942. Venter, Z.S., Cramer, M.D., Hawkins, H.J., 2018. Drivers of woody plant encroachment over Africa. Nat. Commun. 9, 2272. Vikhamar, D., Solberg, R., 2003. Subpixel mapping of snow cover in forests by optical remote sensing. Remote Sens. Environ. 84, 69–82. Volder, A., Briske, D.D., Tjoelker, M.G., 2013. Climate warming and precipitation redistribution modify tree-grass interactions and tree species establishment in a warmtemperate savanna. Glob. Chang. Biol. 19, 843–857. Wang, J., Xiao, X., Qin, Y., Dong, J., Geissler, G., Zhang, G., Cejda, N., Alikhani, B., Doughty, R.B., 2017. Mapping the dynamics of eastern redcedar encroachment into grasslands during 1984–2010 through PALSAR and time series Landsat images. Remote Sens. Environ. 190, 233–246. Wang, J., Xiao, X., Qin, Y., Doughty, R.B., Dong, J., Zou, Z., 2018. Characterizing the encroachment of juniper forests into sub-humid and semi-arid prairies from 1984 to 2010 using PALSAR and Landsat data. Remote Sens. Environ. 205, 166–179. Weisberg, P.J., Lingua, E., Pillai, R.B., 2007. Spatial patterns of pinyon–juniper woodland expansion in Central Nevada. Rangel. Ecol. Manag. 60, 115–124. Wilson, J., Schmidt, T., 1990. Controlling eastern redcedar on rangelands and pastures. Rangelands 12, 156–158. Wimberly, M.C., Janssen, L.L., Hennessy, D.A., Luri, M., Chowdhury, N.M., Feng, H., 2017. Cropland expansion and grassland loss in the eastern Dakotas: new insights from a farm-level survey. Land Use Policy 63, 160–173. Wimberly, M.C., Narem, D.M., Bauman, P.J., Carlson, B.T., Ahlering, M.A., 2018. Grassland connectivity in fragmented agricultural landscapes of the north-central United States. Biol. Conserv. 217, 121–130. Wolter, P.T., Townsend, P.A., Sturtevant, B.R., Kingdon, C.C., 2008. Remote sensing of the distribution and abundance of host species for spruce budworm in Northern Minnesota and Ontario. Remote Sens. Environ. 112, 3971–3982. Wolter, P.T., Berkley, E.A., Peckham, S.D., Singh, A., Townsend, P.A., 2012. Exploiting tree shadows on snow for estimating forest basal area using Landsat data. Remote Sens. Environ. 121, 69–79. Wright, C.K., Wimberly, M.C., 2013. Recent land use change in the Western Corn Belt threatens grasslands and wetlands. Proc. Natl. Acad. Sci. U.S.A. 110, 4134–4139. Wulder, M.A., Dymond, C.C., White, J.C., Leckie, D.G., Carroll, A.L., 2006. Surveying mountain pine beetle damage of forests: a review of remote sensing opportunities. For. Ecol. Manage. 221, 27–41. Wylie, B.K., Meyer, D.J., Choate, M.J., Vierling, L., Kozak, P.K., Green, R.O., 2000. Mapping woody vegetation and eastern red cedar in the Nebraska Sand Hills using AVIRIS. AVIRIS Airborne Geoscience Workshop. JPL Publication, Pasadena, CA, pp. 00–18. Yang, J., Weisberg, P.J., Bristow, N.A., 2012. Landsat remote sensing approaches for monitoring long-term tree cover dynamics in semi-arid woodlands: comparison of vegetation indices and spectral mixture analysis. Remote Sens. Environ. 119, 62–73. Zavaleta, E., 2000. The economic value of controlling an invasive shrub. Ambio 29, 462–467.
9