Trends in Lesser Prairie-Chicken Habitat Extent and Distribution on the Southern High Plains

Abstract

The lesser prairie-chicken (Tympanuchus pallidicinctus) is a species of prairie grouse that occupies grassland ecosystems in the Southern and Central High Plains of the Great Plains. Reduced abundance and occupied ranges have led to increased conservation efforts throughout the species’ range. Habitat loss is considered the predominant cause of these declines. In the Southern High Plains of Texas and New Mexico, lesser prairie-chicken habitat corresponds to the Sand Shinnery Oak Prairie Ecoregion, which is comprised of a mixture of sand shinnery oak (Quercus havardii)-dominated grasslands, sand sagebrush (Artemisia filifolia)-dominated grasslands, and mixed grasslands. In sand shinnery oak–grassland communities, conversion to row-crop agriculture, continuous unmanaged livestock grazing, restriction of natural fire, invasive plant species (e.g., mesquite (Prosopis spp.)), extensive use of herbicides, energy development, and a variety of other factors have also negatively affected ecosystem extent and function. We integrated historical maps and remote sensing-derived information to measure trends in the extent and geographical distribution of sand shinnery oak prairies in eastern New Mexico and northwest Texas. Potential lesser prairie-chicken habitat was reduced by 56% from a potential of 43,258 km² to 18,908 km² in ~115 years (since pre-settlement). Our assessment indicated both mixed grasslands and sand shinnery oak-dominated grasslands were transformed from large parcels of existing vegetation communities to urban settlements, row crops, roads, and industrial land uses by the 1970s. Currently, potential habitat is highly fragmented and restricted to isolated locations in Texas and New Mexico, with an increasing dominance in mixed grasslands, especially in the southeastern portion of the lesser prairie-chicken range. Sand shinnery oak-dominated grasslands have been declining rapidly, from 69% of its potential extent in 1985, 65% in 1995, 54% in 2005, to 42% in 2015. Mixed grasslands drastically declined to 50% of its potential distribution by 1985. Since then, it has been stable until the 2005–2015 period when it declined to 45% of its potential extent. Based on the 2015 assessment, the current potential habitat for lesser prairie-chicken is estimated at 18,908 km² (1,890,800 ha or 4.6 million acres), where 13,126 km² corresponds to mixed grasslands and 5782 km² corresponds to sand shinnery oak-dominated grasslands.

1. Introduction

The Sand Shinnery Oak Prairie (SSOP) ecosystem encompasses a region of sand hills, dunes, and flat sandy aquifer recharge areas in western Oklahoma, northern and western Texas, and eastern New Mexico, with scattered fragments also in southern Utah and adjacent Arizona ([1]; see Figure 1A). The SSOP is characterized by the presence of sand shinnery oak (Quercus havardii) occurring within a matrix of vast mixed grass communities. Sand shinnery oak is a deciduous, long-lived, short shrub, which is typically 1–2 m tall [2] and has an extensive root and rhizome network, as well as numerous physiological and morphological aboveground adaptations that makes the plant extremely drought and fire tolerant [1,3].

Sand shinnery oak prairie ecosystems formed from relic and ancient sands deposited from the Permian Sea and rivers between 4000 and 7000 years ago [4]. Historically (i.e., before anthropogenic land conversion), SSOP composed ~5000–10,000 km² of the semi-arid southwest [1,5]. However, the amount of SSOP on the southern high plains (SHP) has declined, primarily due to conversion to other land-cover types. The presence and spread of invasive species (e.g., mesquite (Prosopis spp.)), expansion of sand shinnery oak monocultures, misuse of the herbicide tebuthiuron, and unmanaged livestock grazing, as well as oil, gas, and wind energy development have also affected the quantity and quality of SSOP on the SHP. The landscape in the ecoregion currently bears little resemblance to pre-European colonization [6,7,8,9].

Currently, sand shinnery oak occurs in combination with shrubs, forbs, and nondominant grasses, including mid- to tall-stature grasses (e.g., big bluestem; Andropogon gerardii) not normally found in semi-arid regions [10]. Changes to landscape structure, as well as the geographical distribution of vegetation types, have been of such frequency and magnitude that scientific consensus is lacking regarding the historical (i.e., pre-settlement) vegetation composition and structure of the SHP, including the extent of SSOP.

Fauna associated with plant communities within SSOP are unique. Of primary interest in this ecosystem is the lesser prairie-chicken (Tympanuchus pallidicinctus), which is a species of conservation concern previously listed as a threatened species under the 1973 Endangered Species Act in 2014 due to the continued reduction in its historical range and population declines associated with loss, degradation, and fragmentation of its habitat [11,12]. Currently the US Fish and Wildlife Service announced its intent to list the species as endangered under the ESA in the SSOP.

Lesser prairie-chickens (LPC) are managed in four discreet ecoregions across their distribution: (1) Short-Grass Prairie/Conservation Reserve Program (CRP); (2) Sand Sagebrush Prairie; (3) Mixed-Grass Prairie, and (4) SSOP (from the Lesser Prairie-Chicken Interstate Working Group Range-Wide Management Plan (RWP) [13,14]). LPC require sand shinnery oak to persist in the SSOP ecoregion [15,16,17,18,19,20,21]. Lesser prairie-chickens lek, nest, and raise broods in sand shinnery oak habitats [18,20,22,23,24,25]. LPC select residual grasses in SSOP for nesting [20] but will nest exclusively in sand shinnery oak in drought years or when residual grasses are not available (e.g., in monocultures of sand shinnery oak; [18]). Nests located in residual grasses are typically surrounded by sand shinnery oak as protective overstory cover [22,23,24]. Deciduous sand shinnery oak leaves are also an important component of nest site selection because they reduce the amount of bare ground surrounding the nest [18]. Brood locations in SSOP are characterized by tall grasses and shrubs (~29–31 cm), including sand shinnery oak.

The use of herbicides [4], particularly tebuthiuron, has been extensive and applied at varying rates since the early 1970s in an effort to increase the relative amount of grass on the landscape by reducing or eliminating sand shinnery oak [19,26,27,28,29,30,31,32]. Although excessive application rates are atypical now, areas originally treated to completely remove sand shinnery oak remain devoid of the plant because clonal rhizomes are lacking, and seedlings from planted acorns do not survive well in shade [33].

Despite the on-going efforts to restore sand shinnery oak communities, uncertainty regarding historical plant community composition and extent is currently hindering widespread implementation of efforts to restore SSOP to benefit LPC populations on these landscapes.

Remote Sensing of LPC Habitat in the SSOP

Remote sensing plays an important part in the generation of historical and contemporary geographical information for endangered species [34]. Time series of satellite imagery have been widely used for understanding the dynamics of species habitats in the US, but very little information about the extent, distribution, and composition of LPC habitat in the SHP portion of the SSOP ecoregion exists. Most of the information available only provide overviews of SSOP distribution from global and national ecoregion maps. Detailed information about habitat patches distribution and composition is lacking. Given that the changes in vegetation in the SSOP can be traced back to the late 1800s, the generation of historical and contemporary habitat information requires methodological approaches that need to combine multiple sources of data, including physical maps, aerial photography, and time series of satellite imagery. The goal of this study is to combine geographical data from multiple sources for reconstructing historical LPC habitat extent and distribution and quantifying the changes it has undergone since then.

Quantifying changes in historical extent and the relative composition of SSOP is possible by integrating pre-settlement vegetation maps, aerial photography, and remote sensing at discrete time steps into geographic information systems (GIS). Pre-settlement vegetation maps produced by academic institutions and federal and state agencies are available in several map libraries and GIS repositories in Texas and New Mexico, as well as digital archives managed by the U.S. Geological Survey (USGS) and Texas Natural Resources Information System (TNRIS). Aerial photography, starting in the 1930s, is readily available from digital repositories at TNRIS and USGS. These data repositories facilitate mapping the historical distribution of vegetation associations on the SHP throughout the 1970s. Multispectral datasets from the Landsat program, on the other hand, are able to provide frequent observations of land characteristics, which can be used for understanding the contemporary LPC habitat distribution and temporal dynamics.

Our goal was to quantify the extent and geographical distribution of SSOP on the SHP to provide information about the percent loss of this plant community within the range of the lesser prairie-chicken. We used spatial data sources from the 1830s to 1930s to describe the pre-settlement extent of vegetation communities of the SSOP on the SHP; then, we evaluated changes in SSOP extent, composition, and land cover from pre-settlement conditions to the present day. This allowed us to derive the baseline information needed to understand how changes in extent, composition, and distribution of SSOP have shaped lesser prairie-chicken populations within the SSOP ecoregion (sensu [14]), which have not been assessed to date.

Our study aims to contribute to the science of historical habitat mapping by providing a methodological approach for generating a coherent time series of habitat information in spite of the challenges of using multiple sources of geographical data. We hope the methods and analytical framework used in this study can be used as a reference for other studies on historical wildlife habitat distribution and composition mapping.

2. Methods

2.1. Study Area

The spatial domain of the study area was the putative LPC historical range in the New Mexico–Texas border area ranging from 32°N to 35.5°N and 104.5°W to 101.5°W (Figure 1A; www.lpcinitiative.org (accessed on 15 November 2021)). Because sand shinnery oak prairies are restricted to sandy soil [19], our study area was limited to the distribution of sandy soils. We used the digital layer of the US Department of Agriculture Soil Survey Geographic Database (USDA SSURGO) [34] to select and define the geographical extent of the sandy soils in the area (see Figure 1B). Within this region, we targeted three vegetation associations, as defined by [20]: (a) sand shinnery oak prairie, which corresponded to land dominated by sand shinnery oak with interspersed sand sagebrush and grasses that included, but was not limited to, sand bluestem, little bluestem, Indiangrass, switchgrass, and sand dropseed; (b) mixed-grass prairies, which included grasses such as little bluestem and western wheatgrass, where sand sagebrush and sand shinnery oak were relatively evenly distributed among a high diversity of shrubs, grasses, and forbs; and (c) sand sagebrush, which corresponded to land dominated by sand sagebrush associated with sand bluestem, grama grasses, sand reedgrass, little bluestem, sand dropseed, and numerous forbs.

2.2. Pre-Settlement to 1970s

One of the goals of our analysis was to create a pre-settlement map of the SSOP distribution as well as creating maps derived from the best available data sources between pre-settlement and the advent of satellite imagery in the 1970s. To assess the pre-settlement vegetation distribution of the SSOP, we surveyed potential sources of geospatial data on vegetation distribution prior to 1970. We contacted and visited the Texas and New Mexico General Land Office Archive, Southwest Collection at Texas Tech University, Geospatial Technologies Laboratory Data Archive at Texas Tech University, University of New Mexico Libraries, Earth Data Archive Center (EDAC) at the University of New Mexico, University of Texas—Austin Libraries, Texas Natural Resources Information System (TNRIS), and USGS and Texas Parks and Wildlife Department GIS digital archives. At each of these physical and online repositories, we performed an exhaustive search of literature, posters, and maps that offered information on the geographical distribution of SSOP in Texas and New Mexico. Maps found in physical format were scanned and digitized into vector format through on-screen digitization using ArcGIS Desktop (ESRI). Given that some maps were originally created at coarse geographic scales (>1:1,000,000), we refined potential SSOP distributions by intersecting them with the distribution of sandy soil as defined by the USDA SSURGO database [35]. The goal of this analysis was first to create a pre-settlement map of SSOP distribution and, secondly, create maps from the best available data sources between pre-settlement and the advent of satellite imagery in the 1970s.

We located seven pre-1970s and contemporary maps in physical and digital format that were available at the University of New Mexico Libraries, EDAC, and the Natural Heritage of New Mexico (NHNM). For New Mexico, the maps depicted the distribution of vegetation cover types for the years 1880, 1957, 1974, 1977, 1978, 1993, and 2011. We did not find pre-settlement or pre-1970s vegetation maps for Texas. The earliest map of vegetation cover types for Texas was available for 1984 in digital format from Texas Parks and Wildlife Department (TPWD), followed by an updated assessment available with information for the years 2005/2006. However, the only source of pre-settlement extent available for Texas could be estimated from the ‘Ecoregion of Texas’ map by [36] who delineated ecoregions at 1:2,500,000 for the US Terrestrial Ecoregions dataset coordinated by the USGS. A map of ecoregions for New Mexico from the US Terrestrial Ecoregions dataset was also available [37]. For consistency, we then merged the Texas and New Mexico Ecoregions maps by [36,37] to show the most coherent delineation of potential sand shinnery oak and grassland extent for both states for pre-settlement conditions.

For comparison of the pre-settlement conditions to a more recent SSOP distribution, we merged a vegetation distribution map from [8] for New Mexico and a vegetation distribution map from [7] for Texas to generate a map of the sand shinnery oak-dominated and mixed grassland-dominated cover types for both states. These maps were published 10 years apart (1984–1993) but represent contemporary conditions of vegetation (circa 1980) that can be compared to pre-settlement conditions. Both maps were generated using similar techniques (aerial imagery and remote sensing data interpretation and manual digitization). The Dick-Peddie/TPWD map was further limited to the potential distribution to sandy soils as defined by the USDA SSURGO database [35].

2.3. Remote Sensing Data Collection (1985–2015)

We used Landsat 5 Thematic Mapper (TM) Collection 1 surface reflectance imagery and Landsat 8 Operational Land Imager and Thermal InfraRed Sensor (OLI/TIRS) Collection 1 surface reflectance imagery to create a coherent time series of satellite imagery in 10-year time steps: 1985, 1995, 2005, and 2015 (

Table 1. Temporal coverage of Landsat imagery used for mapping the extent and loss of the Sand Shinnery Oak Prairie Ecoregion, New Mexico and Texas Southern High Plains, 1985–2015.

Data Source	Year		Date Range for Imagery Collections		No. of Images		Bands
			Winter	Summer	Winter	Summer
Landsat 5	1985	Start date:	1 December 1985	1 June 1986	57	86	B1–B5, B7, cloud mask (cfmask)
		End date:	31 March 1986	30 September 1986
	1995	Start date:	1 December 1994	1 June 1995	80	78
		End date:	31 March 1995	30 September 1995
	2005	Start date:	1 December 2004	1 June 1976	73	80
		End date:	31 March 2005	30 September 2005
Landsat 8	2015	Start date:	1 December 2014	1 June 2015	69	93	B1–B7, cloudmask (cfmask)
		End date:	31 March 2015	30 September 2015

). We acquired imagery to cover all paths/rows (WRS-1) that represented the entire study area. To enhance spectral separability between land cover and land-use classes, we obtained winter and summer season imagery for each time step. We defined winter and summer seasons as two, separate four-month periods in each of the time steps (Table 1). To optimize satellite imagery acquisition process, we created mosaicked image composites circa 1985, 1995, 2005, and 2015 in Google Earth Engine (GEE), a cloud computing platform providing large archives of earth observation data as well as computation capability for efficient data processing and analysis [38].

2.4. Satellite Image Processing and Classification

We used GEE for satellite imagery acquisition, cloud coverage removal, and data exporting via customized JavaScript codes in the GEE code editor platform. We used the cfmask band to ensure all pixels were clear of cloud contamination. We applied a function to calculate the median values for single pixels based on data available in each season. This function was applied to all 711 cloud-free images (total of bands in all scenes) to calculate the median values of all stacked images, at each pixel for each year and for each season and band. This process allowed to create a cloud-free image composite for each season and time step. We exported processed imagery to Google Drive, and then downloaded these images to our local computer for further processing and analysis. Given that SSOP is strictly limited to the distribution of sandy soils, satellite imagery was then subset to the extent of sandy soil coverage as defined by the USDA SSURGO geospatial database (Figure 1B).

For each time step of the time series, we applied a semi-automated classification approach, which included a machine-learning image classification process with posterior class recoding and spatial filtering. For the supervised classification, imagery for each time step included all optical bands (winter/summer) plus three vegetation indices: the normalized difference vegetation index (NDVI), the enhanced vegetation index (EVI), and simple ratio (SR) [39,40]. We used vegetation indices as additional data to expand the capability of the machine learning algorithm for optimal variable selection when constructing decision trees.

We conducted field verification of land covers in the summer of 2016 in selected locations within the study area. The purpose of the field verification was to gain in situ knowledge of the plant community spectral and textural characteristics as observed in high-resolution satellite imagery (<1 m pixel size) in Google Earth Pro and Landsat imagery. We randomly selected locations within areas where we had access to (through public roads and where in situ verification in private lands was possible). We inspected land nearby public roads in southern and central areas of the SSOP distribution (Ward and Yoakum counties) and geolocated 178 locations of sand shinnery oak-dominated, sand sagebrush-dominated, or mixed grasslands. Field data collection for image interpretation was supported by very-high-resolution imagery collected using a DJI Phantom 3 Unmanned Aerial Vehicle (UAV) over representative areas of sand shinnery oak-dominated and grassland-dominated patches in the Weaver Ranch in Roosevelt County, New Mexico (Figure 2). We used UAV imagery to visually identify different vegetation types and compare to Landsat false-color composites spectral signatures.

We created training samples by interpreting high-resolution imagery in Google Earth Pro with the support of field data to aid in satellite image classification (Figure 2). We used six classes of training samples, consisting of sand shinnery oak, sand sagebrush, grasslands, urban area, water, and croplands. Finally, we created 1100 training samples for each class for each timestep of the time series.

Training samples were used to train a random forest machine learning algorithm [41] implemented in ArcGIS 10.5 (ESRI) for image classification (e.g., “Random Trees algorithm”). We used the default parameters of the algorithm for maximum tree depth (n = 30) but adjusted the maximum number of trees allowed to 500 to maximize the number of possible trees evaluated. During preliminary classification attempts, results greatly improved when spectrally similar classes were merged; therefore, we unified the training sites to represent four major classes: sand shinnery oak-dominated grasslands, mixed grasslands (including sand sagebrush), water, and anthropogenic land uses (e.g., roads, buildings, urban and rural developed land, fallow agricultural land, and row crops with center pivot irrigation).

Preliminary sand shinnery oak-dominated grasslands and mixed grasslands maps were spatially coherent with the distribution of the same target land covers as high-resolution imagery in 1985 and 2015. For 1995 and 2005, however, preliminary classification results showed large mapping omission and commission errors, probably associated with soil moisture and vegetation phenology differences with respect to 1985 and 2015, which might have affected the accuracy of the collected training samples. To avoid error propagation from the wall-to-wall comparison of the hard-image classifications [42], we applied the pairwise image differencing method to classify SSOP vegetation in 1995 and 2005 [43]. For the application of the image differencing method, we isolated the sand shinnery oak-dominated grasslands class in the 1985 map as a single vector layer and explored the subsequent negative and positive changes that occurred in the Modified Soil Adjusted Vegetation Index (MSAVI, [44]) values derived from the 1985 and 1995 summer imagery.

By exploring different thresholds of change in MSAVI values within the sand shinnery oak class and comparing to contemporary high-resolution imagery in Google Earth Pro, we were able to identify losses in sand shinnery oak cover as well as gains in grassland and anthropogenic land from 1985 to 1995 and apply these changes to the 1985 map to create a final 1995 map. The same process was repeated for the mixed grassland and anthropogenic classes. This manual process was more effective for mapping the distribution of sand shinnery oak, grasslands, and anthropogenic land-cover classes, as observed in the 1995 and 2005 imagery.

Final maps were visually inspected for large classification errors by comparing them to Landsat Imagery and high-resolution imagery in Google Earth Pro. Pixel recoding was performed when possible. Although there were no hard criteria on the selection of large errors, the detection of errors by visual inspection of the map allowed us to make sure that erroneous polygons with an approximate minimum size of 50 hectares were eliminated or recoded to the correct class. The maps were further converted to vector format and limited to class polygons of area > 0.01 km². This reduced the effect of including misclassified single pixels in the final product.

In addition, a set of independent validation points for sand shinnery oak-dominated and mixed grasslands land-cover types was collected by a trained GIS analyst to estimate the accuracy of the land-cover maps produced by automated classification. For assessing the accuracy of the 2015 map, the analyst collected 1801 validation points in geographical locations with continuous coverage of sand shinnery oak-dominated land and mixed grasslands using high-resolution imagery available in Google Earth Pro (Figure 2). For assessing the accuracy of the 1985 map, 967 validation points were collected for the same vegetation types. We compared the validation points to the classified pixels and calculated accuracy metrics following standard methods by [45] for creating a simple confusion matrix and deriving the overall user’s and producer’s accuracy measures.

In addition, we applied the methodology proposed by [46] for calculating class quantity difference and exchange for the remotely sensed time series of maps (1985–2015). This was applied in order to identify patterns of change between shinnery oak and mixed grasslands and assess whether the sources of these changes were caused by remote sensing errors or land system dynamics. For assessing the accuracy of the mapped loss patterns, we selected 150 random locations where shinnery and mixed grassland areas were mapped as stable (persistent) and 150 random locations where vegetation was converted to anthropogenic land cover (loss). We inspected each point using historical high-resolution imagery available in Google Earth Pro and generated an error or confusion matrix [45].

2.5. Trends in Landscape Composition

For assessing changes in the relative proportion of sand shinnery oak-dominated and mixed grassland land covers, we performed two independent analyses at different temporal and spatial scales. The first analysis provides a quantification of change in 21 selected locations over a time period spanning up to 75 years. The objective of this analysis was to identify broad land-cover and land-use changes observed in a time period close to the historical record. The second analysis provides a spatial quantification of change in 10-year time steps from 1985 until 2015 (30-year time span). The objective of this analysis is to provide a finer resolution quantification of changes in land cover and land use over the most recent decades.

In the first analysis, we compared the earliest available remote sensing information (aerial photographs) to the most recent classified map of 2015. Early aerial photo availability was sparse within the study area; therefore, we selected locations where aerial photos were available for at least two dates between 1930s (first flight missions for the region) and 1970s to compare to the Landsat-based maps for the 1985–2015 time series. We performed an exhaustive search on the hundreds of aerial photo mosaics available through the USGS Earth Explorer data repository (http://earthexplorer.usgs.gov/ (accessed on 1 September 2017)). We acquired 121 aerial photo mosaics for 21 selected locations. Aerial photos were georeferenced, mosaicked, and interpreted at the 1:25,000 geographic scale. The distribution of sand shinnery oak-dominated and mixed grasslands in aerial photos was manually digitized in ArcGIS.

Using data interpreted from aerial photos (circa 1950s) and data from the 1985–2015 time series, we quantified the total change suffered by shinnery oak and mixed grasslands over a 75 years time span within the 21 selected locations where aerial photos were available. Instead of performing this analysis for the entire aerial photo mosaic coverage at each location, the analysis was performed using a grid with a cell size of 2.5 km × 2.5 km over each selected location in order to explore spatial patterns of change. For visualization of trends, we used six broad categories to show the level of positive change and levels of negative change.

In the second analysis, we used a 10 km × 10 km grid over the whole study area for calculating changes in vegetation cover for each class and each time step during the period 1985–2015 using classified maps generated in the previous sections. We generated a heatmap showing a class for negative change in proportion greater than 20% and a class for negative change in proportions between 10 and 20%. We did the same for positive changes of 10–20% and >20% expansion. Changes in the ranges between −10% and +10% were considered relatively stable. Given that a sustainable population of lesser prairie-chickens requires approximately ~100 km² of contiguous high-quality native grassland [14], we also assessed the change in the number of fragments and the mean patch size of both vegetation categories as well as the change in the proportion of fragments in the <100 km² and >100 km² size classes for the 1985–2015 period.

3. Results

3.1. Pre-Settlement to 1970s Vegetation Maps

The comparison between the pre-settlement and 1984/93 map composite shows a clear pattern of decline in extent of sand shinnery oak-dominated grassland cover types from pre-settlement to the 1980s in response to the expansion of anthropogenic land uses in west Texas and eastern New Mexico (Figure 3). The potential extent of sand shinnery oak-dominated grasslands prior to settlement was estimated at 13,801 km², while mixed grasslands extended for 29,457 km², which totaled 43,258 km² (~10 million acres) of potential habitat for the lesser prairie-chicken. Circa 1980s, the extent of sand shinnery oak-dominated grassland decreased to 10,959 km² (−21%) while mixed grasslands decreased to 13,965 km² (−53%). The majority of habitat loss occurred in Texas (73% decrease in grassland area and 15% decrease in sand shinnery oak area).

3.2. Landsat-Based Vegetation Mapping

Using time series of Landsat TM and OLI data for 1985, 1995, 2005, and 2015, we mapped the change in extent and distribution of sand shinnery oak-dominated grasslands and mixed grasslands in the SHP within the lesser prairie-chicken population range. An accuracy assessment implemented using 1801 validation points for the 2015 vegetation map indicated an overall accuracy of 81% for all classes, with a user’s accuracy of 80% for the mixed grasslands class and 82% for the sand shinnery oak class. Producer’s accuracy for mixed grasslands was 85% and 76% for the sand shinnery oak class. Similarly, an accuracy assessment for the 1985 map was implemented using 967 validation points, indicating an overall accuracy of 86% for all classes, a user’s accuracy of 87% for mixed grasslands, and 84% for the sand shinnery oak class. Producer’s accuracy for both classes were 90% and 80%, respectively, for the 1985 map. The error matrix calculated based on 300 reference loss/no-loss occurrences during the 1985–2015 time interval showed an overall accuracy of 71% for the 1985–2015 change map, with a producer’s and user’s accuracy for the loss category of 77% and 58%, respectively.

Figure 4A shows that overall annual change was less during the 1985–1995 and 1995–2005 periods and greater during 2005–2015. The size of each rectangle in Figure 4A is proportional to the size of the change during the time interval. Most of the change is a quantity difference in the first two periods with little exchange. During 2005–2015, the quantity difference and exchange components were similar. The graph describes that the conversion between shinnery oak to grasslands was higher during the 2005–2015 period. Through visual interpretation of exchange patterns, especially prominent in the south and western portions of the map, we observed that rather than remote sensing errors and spectral class confusion, exchange between shinnery oak and grassland did occur during the 2005–2015 period. This exchange was less prominent during the 1985–2005 period. The overall patterns of gain and loss during the 1985–2015 period are shown in Figure 4B.

When compared to the 1993/1984 map composite based on the Dick-Peddie/TPWD maps, the satellite image-based estimates for 1985 were slightly lower for sand shinnery oak (a difference of 1374 km²) and slightly greater for grasslands (a difference of 822 km²;

Table 2. Summary statistics for the 1985–2015 Landsat-based assessment to map vegetation communities that support lesser prairie-chicken populations on the New Mexico and Texas Southern High Plains.

	1985		1995		2005		2015
Class	km2	% Cover from Total Potential Distribution	km2	% Cover from Total Potential Distribution	km2	% Cover from Total Potential Distribution	km2	% Cover from Total Potential Distribution
Sand Shinnery Oak	9585	69.45	9011	65.29	7397	53.59	5782	41.89
Mixed Grasslands	14,787	50.19	15,218	51.66	15,952	54.15	13,126	44.55
Total	24,373	56	24,229	56	23,349	54	18,908	44

, Figure 3 and Figure 5). This slight difference was expected given the methodological differences for achieving the same result; however, both maps were spatially coherent and showed spatial agreement for a similar time period.

3.3. Trends in Landscape-Level Composition

For the first analysis spanning up to 75-years, time series of aerial photo mosaics available between the 1930s and 1970s corresponded in majority to aerial photos from 1953 and 1954, and photos from 1967 to 1978. Results from the interpretation of aerial photography from these periods were compared to the 2015 Landsat-based classification using a 2.5 km × 2.5 km grid in 21 selected locations across the study area. Moderate increasing/decreasing trends in shinnery oak cover are prevalent from the 1950s until 2015; however, a few clusters of grid cells with a steep reduction in shinnery oak cover were found in Terry, Yoakum, Gaines, and Andrews counties. A cluster of cells with a steep increase in sand shinnery oak cover was registered only for Bailey County.

In the second change analysis based on the recent 1985–2015 classified maps, spatial patterns show a tendency for expansion of mixed grasslands since 1985 in the majority of cells in the southern portion of the study, but also consistent declines in the central to northern portions of the study area. Trends in mixed grassland expansion have been consistent after 2005, especially in the southern portion of the study area (Lea, Andrews, and Gaines counties). In contrast, sand shinnery oak-dominated grasslands declined in the majority of the cells in their entire range, especially since 1995.

In general, Figure 6 and Figure 7 show a similar historical trend, where both vegetation types have consistently declined since pre-settlement and mid-century conditions due to anthropogenic expansion, especially in the southern portion of the study area. However, as observed in Figure 8, spatial and temporal patterns of decline and expansion are variable across the region in recent times.

Land-cover distribution metrics show a dissimilar pattern of change in patch configuration for the sand shinnery oak-dominated grasslands and mixed grasslands (

Table 3. Summary of the fragmentation statistics from the 1985–2015 Landsat-based assessment to map vegetation communities that support lesser prairie-chicken populations on the New Mexico and Texas Southern High Plains.

	1985			2015
Class	Number of Fragments	Mean Patch Size (km2)	% of Total Area	Number of Fragments	Mean Patch Size (km2)	% of Total Area
Sand Shinnery Oak
0 km2–0.1 km2	11,081	0.02	3.09	8263	0.02	3.81
0.1 km2–1 km2	1372	0.26	3.73	995	0.27	4.76
1 km2–10 km2	145	2.22	3.36	134	2.62	6.09
10 km2–100 km2	9	19	1.79	31	34	18.54
>100 km2	15	562	88.02	10	293	66.81
Mixed Grasslands
0 km2–0.1 km2	10,864	0.02	1.87	9011	0.02	1.61
0.1 km2–1 km2	1133	0.25	1.99	551	0.25	1.08
1 km2–10 km2	169	3.11	3.56	57	2.57	1.12
10 km2–100 km2	77	47	24.61	16	44	5.38
>100 km2	33	304	67.98	45	563	90.80
All classes
0 km2–0.1 km2	21,945	0.02	2.35	17,274	0.02	2.28
0.1 km2–1 km2	2505	0.25	2.67	1546	0.27	2.21
1 km2–10 km2	314	2.70	3.48	191	2.61	2.64
10 km2–100 km2	86	44	15.63	47	37	9.41
>100 km2	48	385	75.86	55	514	83.46

). For sand shinnery oak-dominated grasslands, 88% of its extent was distributed in large patches of >100 km² by 1985; however, that decreased to 66% by 2015. The number of smaller fragments has declined with the exception of mid-size fragments between 10 and 100 km², which increased. For mixed grasslands, although there has been a general reduction in its extent (Figure 6 and Figure 8), the current (2015) patch configuration shows a considerable increase in large patch sizes (>100 km²) since 1985, while the number of small and mid-size fragments has declined (Table 2 and Table 3). Loss was spatially and temporally dynamic (Figure 8).

Results from the 1985–2015 Landsat-based time series assessment indicate that lesser prairie-chicken habitat was reduced to 56% of its potential distribution in the SHP by 1985; thereafter, it decreased to 54% in 2005 and 44% by 2015 (Figure 6). However, sand shinnery oak and mixed grasslands have shown dissimilar patterns in their reduction. Sand shinnery oak-dominated grasslands declined rapidly to 69% of its extent by 1985, then to 65% in 1995, then 54% in 2005, and finally 42% in 2015. Mixed grasslands declined to 50% of its distribution by 1985. Since 1985, our results suggest that the extent of mixed grasslands were stable to slightly decreasing through the 2005–2015 period. By 2015, mixed grasslands were reduced to 45% of its extent. Based on the 2015 map, the potential habitat available for the lesser prairie-chicken was estimated as 18,908 km² (1,890,800 ha or 4.6 million acres), where 13,126 km² corresponds to mixed grasslands and 5782 km² corresponds to sand shinnery oak-dominated grasslands in the SSOP on the SHP.

4. Discussion

4.1. Implications of LPC Habitat Conservation

Using multiple sources of geographical data, we quantified a clear, declining trend in the extent of potential LPC habitat in the SSOP on the SHP from 1851 to 2015. Our results indicated that sand shinnery oak prairies and mixed grassland vegetation communities within the SSOP were reduced by ~56% of their combined, potential extent in approximately 115 years. The decline includes losses that have occurred within the last 30 years (1985–2015), especially for sand shinnery oak-dominated grasslands, which are likely correlated to previously reported factors that influence the amount and quality of vegetation communities that support lesser prairie-chickens [19].

Declining availability and quality of lesser prairie-chicken habitat in the Southern High Plains have previously been reported [1,4,20], but estimates of past and present extent and vegetation distribution (within the context of lesser prairie-chicken ecology) differ among studies. Our study is the first to quantify historical changes in the relative proportion of sand shinnery oak prairies and mixed grasslands in the SSOP on the SHP. Our findings are consistent with [47], where the authors found vegetation communities that support lesser prairie-chicken populations in Kansas decreased within the time frame reported herein (~1850–2015). However, most of the land conversion in Kansas occurred by the 1970s, and additional losses between the 1970s to now were mitigated by implementation of CRP [47]. Our data corroborate previous assessments that speculated loss in the extent of vegetation communities, important for facilitating lesser prairie-chicken lekking, nesting, and brood rearing, is the one of the primary mechanisms responsible for long-term population declines in the SSOP ecoregion on the SHP [19]. Contemporary lesser prairie-chicken populations in SSOP were the most abundant in the mid-1980s [48] and declined considerably between the late 1980s and early 1990s. Lesser prairie-chicken densities within the SSOP on the SHP have fluctuated widely since the mid-1990s and have failed to reach similar densities as the ones reported in [48].

Only part of the sand shinnery oak-dominated grasslands were replaced by mixed grasslands during the time frame of our assessment. Our assessment indicated both mixed grasslands and sand shinnery oak-dominated grasslands were transformed from large parcels of existing vegetation communities to urban settlements, row crops, roads, and industrial land uses (mostly oil exploration and extraction) by the 1970s. Our analysis indicated this trend has continued since the 1970s in both vegetation communities, which also corroborates previous studies that the speculated loss in extent of native prairie for anthropogenic uses is one of the primary mechanisms responsible for long-term lesser prairie-chicken population declines in the SSOP. Anthropogenic effects on landscapes and vegetation communities that support lesser prairie-chicken extend beyond outright loss of habitat. Lesser prairie-chickens avoid anthropogenic features when nesting [18,49,50,51]; their home ranges are often constrained within areas without human structures [52,53]; their survival varies among areas with disproportionate anthropogenic development [54,55]; and contemporary evidence suggest functional connectivity may be limited in areas with wind farms and other development [21]. Our assessment adds to the weight of the evidence that anthropogenic development within lesser prairie-chicken ranges corresponds to long-term population declines.

We expected to see expansion of grass-dominated vegetation communities in the SHP given widespread application of tebuthiuron and other herbicides to eliminate sand shinnery oak in favor of grasses for forage production. Previous authors indicated misuse and over application of tebuthiuron has resulted in a considerable reduction in sand shinnery oak in the Southern High Plains [4,5,56,57]. Our results indicated that the continued removal of sand shinnery oak in favor of grasses for production is likely one mechanism responsible for the reduction in population carrying capacity of SSOP for lesser prairie-chicken populations. In the counties of Lea (New Mexico), Andrews, and Gaines (Texas), maps show that grasslands have been expanding in coverage, at least during the last 10 years. The pattern of widespread grassland predominance in 2015 for the southern portion of the study area is visible from airborne high-resolution imagery, as seen on Google Earth Pro.

Because sand shinnery oak provides food, thermal cover, escape cover, and supports reproductive activities of lesser-prairie chickens, its loss may be detrimental [17,18,19,20,21,25,58]. However, further research is needed to investigate the direct and indirect drivers behind the observed widespread expansion of grassland-dominated areas in the SSOP and assess the local and regional positive/negative impacts on lesser prairie-chicken populations. The CRP, particularly native mixes, is beneficial to lesser prairie-chickens for nesting and raising broods in SSOP [59]. Spencer et al. [47] found CRP mitigated against native grassland loss in Kansas, and [60] found greater grassland composition, including CRP, buffers negative impacts of extreme drought, which can cause population-level declines in reproductive output and abundance [19,60,61,62]. Our assessment was not conducted at the spatial scale necessary to identify specific drivers, and management of the species would benefit from a better understanding of the underlying causal mechanisms that led to grassland expansion, particularly the timing, general location, and arrangement of CRP within the SSOP.

Our results also show an overall increase in the isolation and reduction in size of large patches of sand shinnery oak-dominated grasslands. Between 1985 and 2015, mixed grasslands had formed more continuous large patches across the region, especially in the southern locations of the study area, but the addition of grasslands in this section of the SSOP does not specifically imply net benefit for lesser prairie-chicken populations. For example, no data exists that support CRP planted in non-native grasses, such as Lehmann lovegrass (Eragrostis lehmanniana), weeping lovegrass (Eragrostis curvula), or old-world bluestems, as beneficial to lesser prairie-chickens, despite recent and on-going studies that support native-seeded CRP as lesser prairie-chicken habitat [59,60]. The measured changes in the relative proportions of small, mid-size, and large patches suggest isolation and contiguity for mixed grasslands but increasing patchiness for sand shinnery oak-dominated grasslands. These trends do not seem to be concentrated over specific regions. Instead, sand shinnery oak loss has been widespread across its range in the SHP. The decline in the extent of mixed grasslands has occurred since 1985 and mostly within in the northern portion of the study area, and our analysis suggest this loss was due to the expansion of agricultural fields, particularly around the Bailey and Lamb counties, Texas. Harryman et al. [59] found lesser prairie-chickens selected CRP in these counties, but not in the Bailey/Lamb County sandhills, which were within dispersal distance considered achievable for the animals radiomarked in their study [63].

Our results show regional and landscape scale historical changes in vegetation type predominance and reveal trends that could be useful for strategic planning in conservation efforts. For example, data presented herein may be maximized if they are used to recalibrate existing geospatial tools used by industry (e.g., Crucial Habitat Assessment Tool, https://www.wafwachat.org/ (accessed on 1 November 2021)). They can also be used to identify areas to increase the patch sizes of vegetation communities used by lesser prairie-chickens (e.g., CRP) and identify areas for land reclamation no longer used by industry (e.g., removal of unused oil/gas pads). Data here-in can also be used to identify areas to test the framework for propagating sand shinnery oak from wild acorns; to protect the existing core distributions of large patches of SSOP; and to implement conservation strategies that benefit the species there-in (e.g., grazing management, avoid development, etc.; [60]).

4.2. Methodological Limitations and Challenges

Limitations of our study include the scarce availability of aerial photography for the entire SHP region which limited part of our historical analysis on composition to only 21 locations across the study area. The use of this sampling scheme, however, still allowed having a snapshot of the regional variability of changes in SSOP plant community composition. In this context, museum records could be used to further refine the limits of the historic range of the LPC by comparing specimen localities with vegetation maps derived from historical imagery.

Other limitations in the remotely sensed data analysis included possible conditions in soil wetness that could have affected surface reflectance from the Landsat TM scenes in this region in 1995 and 2005, therefore causing an overestimation and/or underestimation vegetation abundance and density in these counties. However, careful application of a continuous change detection approach using pairwise image differencing allowed us to generate spatially and temporal coherent land-cover and land-use distributions.

The analysis of LPC habitat at a plant community scale, rather than as a whole, is important for conservation purposes. In this study, we used a random forest image classification algorithm for classifying multispectral imagery into two plant community classes. The use of UAV imagery and proper in situ verification for the interpretation of high-resolution imagery helped in the derivation of training sites and validation points. Our effort provided a time series that can be used with confidence for habitat analysis at 86% and 81% overall accuracies for the 1985 and 2015 maps and a 71% overall accuracy for the change map. However, we believe new remote sensing studies could improve the separation of shinnery oak, mixed grasslands, and anthropogenic land cover by combining multispectral imagery with radar imagery, LiDAR, and other ancillary data, as well as with analytical techniques such as object-oriented image classification and using deep learning algorithms for image classification.

The data generated in this study can also serve as a starting point for designing efforts using very-high-resolution imagery from the National Agricultural Imagery Program (NAIP). NAIP offers the possibility of assessing changes in vegetation composition at several time steps since the mid-1990s up to the present at a 0.5–5-m pixel resolution. Matching time series of NAIP imagery with locations of known lesser prairie-chicken population reductions or expansions can inform the role of habitat configuration in these processes and may help identify the specific proximate mechanisms responsible for population declines at finer spatial and temporal scales addressed herein.

5. Conclusions

We present the first comprehensive assessment of the changes in extent and distribution of the LPC habitat in the SSOP ecoregion of the SHP using multiple sources of remote sensing data.

The study integrates the analysis of historical vegetation maps and remote sensing imagery (1985–2015) to map historical changes in the extent and distribution of sand shinnery oak-dominated grasslands and mixed grasslands in the SHP. The methodology and analytical framework used in this study can serve as an example for generating a complete time series of habitat maps using diverse digital and non-digital data sources.

Understanding historical and contemporary habitat distribution and composition is an imperative for determining the survival and extinction probabilities of wildlife species of concern. In the case of the LPC habitat, our results demonstrate that vegetation communities that support its populations have been reduced by 56% from a potential of 43,258 km² to 18,908 km² in ~115 years (since pre-settlement). Patchiness and isolation of shinnery oak-dominated land is increasing in Texas and New Mexico, with an increasing dominance in mixed grasslands, especially in the southeastern portion of the lesser prairie-chicken range in SSOP, where lesser prairie-chickens have not been observed within the past decade. The removal of sand shinnery oak for row crops, energy development, and grass production is the main driver behind changes in vegetation communities that support lesser prairie-chickens, but further research is needed to assess the causal mechanisms responsible for changes where conversion is not apparent.