Habitat quantity of Red-cockaded Woodpecker Picoides borealis (Aves: Piciformes: Picidae) in its former historic landscape near the Big Thicket National Preserve, Texas, USA

We quantified pine-forested habitat suitable for Red-cockaded Woodpecker Picoides borealis in the former historic range of the species to assess the potential for possible re-colonization.  We used a remotely-sensed image and geographic information systems (GIS) to create a land-use/land (LU/LC) binary cover map, from which we calculated the habitat suitability index (HSI) based on an estimated home range of 50ha.  A sensitivity analysis revealed the necessity for more data to make an accurate estimate, but our analysis of landscape metrics indicates more than 930ha of suitable habitat patches.  These patches are heavily fragmented and mostly located on private lands.  They can be assessed for understory and herbaceous vegetation and can be restored for possible re-establishment of approximately 18 groups/colonies of Red-cockaded Woodpeckers.


INTRODUCTION
Red-cockaded Woodpecker Picoides borealis is Near Threatened (BirdLife International 2013) and nationally an endangered species (USFWS 2014).The bird is endemic to mature pine forests of the southeastern United States, which once extended from Florida to New Jersey and as far west as Texas, reaching inland to Oklahoma, Missouri, Kentucky and Tennessee (Ligon 1970;Jackson 1971;Ferral 1998;USFWS 2005).During the early 19 th century the wide spread of agriculture and timber harvesting led to severe habitat degradation and substantially reduced the woodpecker habitat range, which is currently scattered north from Florida to Virginia and west to southeast Oklahoma and southeastern Texas.The species is no longer found in New Jersey, Maryland, Tennessee, Missouri and Kentucky, while in southeastern Texas birds are mostly found in national forests of Angelina, David Crockett, Sabine and Sam Houston, but not in Big Thicket National Preserve (Conner & Rudolph 1995).The drastic reduction of mature pine forests coupled with modern forestry practices such as a reduced timber-rotation period and fire suppression proved detrimental to woodpecker populations, and the species was listed as endangered in 1973 (Hooper et al. 1980;Conner & Rudolph 1989, 1991, 1995;Costa & Walker 1995).
Red-cockaded Woodpeckers are habitat specialists that require large, old and living species of Longleaf Pinus palustris, Shortleaf P. echinata, Loblolly P. taeda, Pond P. serotina and Slash P. elliotii pine, preferring Longleaf Pine for nesting and foraging (Hooper et al. 1980;Jackson 1994;Hedrick et al. 1998;Conner et al. 2004).The optimal tree age varies with species, i.e., 80-100 years for loblolly and shortleaf pine and 100-120 years for Longleaf Pine with enough heartwood space to support cavity chambers and little or no mid-story hardwood vegetation (Hooper et al. 1980;Conner et al. 1994;Hedrick et al. 1998).Natural or prescribed fires controlled the mid-story overgrowth for decades and the result was open, park-like mature pine woodlands and savannahs with abundant herbaceous ground cover that provided an ideal habitat for these birds.Besides age, the potential cavity tree has high rates of Redheart Fungal Phellinus pini infection that softens the heartwood and facilitates cavity excavation (Conner et al. 1976(Conner et al. , 1994(Conner et al. , 2004;;Conner & Locke 1982;Hooper 1988;Walters 1990).
A colony or cluster is a collection of two to >12 cavity trees in 5-10 acres (approximately 2-4 ha) of land, and the cavity trees are normally located within a one-mile radius from each other (USFWS 2005).A single colony has two to nine birds, with one breeding pair and the rest helpers.A suitable foraging habitat or territory surrounds a colony and covers an area of 30 to 81 contiguous hectares (75-200 acres) of park-like mature pine stands (Hooper et al. 1982;Jackson 1994).Thus only contiguous open stands of mature longleaf and other pine species with herbaceous ground cover offer high quality habitats for Red-cockaded Woodpeckers (Conner & Rudolph 1991).
Few studies exist on the use of geographic information systems (GIS) and remote sensing to study the habitat of Red-cockaded Woodpeckers.Thomlinson (1993) used GIS, remote sensing and landscape ecology to study ecological characters of suitable pine stands in southeastern Texas.Cox et al. (2001) evaluated GIS methods that were used to assess Red-cockaded Woodpecker habitat and cluster characteristics.Ertep & Lee (1994) used GRASS to facilitate Red-cockaded Woodpecker management at Fort Benning Military Reservation.Another recent study by Santos et al. (2010) reports the use of remote sensing based on hyperspectral imagery to study tree senescence in Redcockaded Woodpecker habitats.They used reflectance properties of the bands to detect senesced pine trees and found Red-cockaded Woodpeckers did not inhabit such trees.We utilized GIS and remote sensing techniques to study the spatial distribution of pine forest in one of the former historical ranges of Red-cockaded Woodpeckers (southeastern Texas) and to assess suitable habitats.We also used habitat suitability index (HSI) models and FRASGSTATS to evaluate or quantify species-habitat relationships.HSI models provide a quantitative measure of the quality of wildlife habitats and can integrate our understanding of wildlife-habitat relationships especially at landscape scales (Larson et al. 2003).In addition, process-oriented and empirical HSI models are commonly used to assess wildlife-habitat relationships (Dettki et al. 2003).Process-oriented models assess plausible causal relationships to provide a general conceptual framework; whereas empirical models analyze data on habitat characteristics collected at specific sites (Thapa et al. 2014).
For this paper we adopted a process-oriented approach to develop a heuristic HSI model for the Red-cockaded Woodpecker.This approach is based on a literature review (U.S. Fish and Wildlife Service's HSI models), field observations (ground-truth) and geographic data obtained from topographic maps (scale 1:24000, USGS).An HSI is based on a set of functional relationships between habitat suitability (expressed as a

Habitat quantity of Picoides borealis
Thapa & Acevedo dimensionless index or score) and habitat requirements (variables).These variables are selected according to their relevance to the organism; for example herbaceous canopy cover, tree canopy cover, tree height, tree age and proximity to water.There is a partial suitability for each variable, which scales from 0 (unsuitable habitat) to 1 (optimum habitat).The overall HSI, which also scales from 0 to 1, is calculated with a formula that represents hypothetical relationships between partial suitability indices.GIS provides a tool to synthesize habitat data derived from remotely sensed sources together with databases of elevation, soil types, land use, and land cover.Thus GIS can be coupled with remote sensing to calculate HSI over relatively large geographic areas, and incorporate landscape variables at multiple spatial scales.We also demonstrate the use of GIS and remote sensing to collect or prepare data for habitat fragmentation study by using software called FRAGSTATS, which is a computational program designed to calculate a wide array of landscape metrics from categorical maps (McGarigal & Marks 1994, 1995;McGarigal 2002).Some of the metrics are commonly used to measure and quantify spatial patchiness in terms of composition (patch types and abundance) and configuration (shape and juxtaposition).These metrics represent the percentage of fragmented habitats, area of largest patch, and-most importantly-the area of remaining potentially suitable habitat (Girvetz et al. 2007).
In this paper, we used aforementioned habitat characteristics and applied remote sensing, GIS and FRAGSTATS techniques to examine abundance, distribution and fragmentation of available pine forest and provide a possible scenario for re-colonization by Red-cockaded Woodpeckers.We have four scientific objectives: (1) to use a Landsat Enhanced Thematic Mapper Plus (ETM+) image to develop a land-use/ land-cover map (Laperriere et al. 1980) ; (2) to develop an heuristic GIS-based HSI model and a map for the woodpecker; (3) to determine the spatial distribution of current potentially suitable habitats; and, (4) to illustrate a general methodology for conservation cartography and spatial analysis that can be adapted to other interior-forest-dwelling avifauna of conservation interest.In addition, we have two policy-oriented objectives: (1) to provide a map of potentially suitable Red-cockaded Woodpecker habitat that may be preserved for (a) existing populations in the region or (b) that may serve as sites for establishing new populations in the region; and (2) to indicate the most important habitat characteristics, such as shape, size, and habitat composition for purposes of proactive Red-cockaded Woodpecker habitat management in the region.

Study Area
The study area is located near the Gulf coastal plains of southeastern Texas between the Trinity River to the west and the Neches River to the east, around the small towns of Kountze, Silsbee, Lumberton and suburbs north of Beaumont that adjoins the 39,338ha Big Thicket National Preserve (BTNP) (30-31 0 N to 94-95 0 W) (Fig. 1).Over the last five decades the landscape surrounding the BTNP has been converted from continuous pine forest to a matrix dominated by agriculture, pasture, timber plantations and exurban and suburban development (Wilcove et al. 1986).As a result the pine forests were converted into small patches isolated by a matrix of agricultural or other developed lands (Callicott et al. 2007).The study area was further subjected to intense oil and gas exploration that continues today.While such activities seem to have minimal effects on breeding, proximity to roads and vehicular movement does affect foraging activities of Red-cockaded Woodpeckers (Charles & Howard 1996).Annual precipitation averages 1350mm (Marks & Harcombe 1981;Callicott et al. 2007) and is uniformly distributed throughout the year, but because of its proximity to the Gulf of Mexico the Big Thicket study area experiences a high frequency of devastating tropical storms and hurricanes.Since 1900, 40 tropical storms and hurricanes have struck the Gulf coast, with Rita in 2005 and Ike in 2008 being the most recent big storms to hit Texas (NOAA 2008).However, these hurricanes did not cause damage in the study area as they did in the surrounding counties and areas especially near Galveston Bay, Harris and Angelina (NOAA 2005;Bainbridge et al. 2011).Nevertheless, hurricanes and other extreme natural disturbances such as severe winter can damage large portions of cavity and foraging trees, thereby affecting breeding populations of Red-cockaded Woodpeckers, which in turn leads to loss of genetic diversity (Reed et al. 1988, Bainbridge et al. 2011).
The vegetation types of the study area can be characterized by both community physiognomy and physiographic position.Forests, savannas, and shrub thickets are normally combined with important trees such as pine, oak, and other hardwoods to characterize community physiognomy while upland, slope, floodplain and flatland indicates the physiographic position of the

Field Data Collection and Vegetation Classification
We collected a total of 287 vegetation GPS points in May-June of 2007.We used a cloud-free Landsat image of 2003 to perform supervised classification procedures to derive final land use and land cover (LU/LC) categories.Supervised methods require the user to define the spectral characteristics of known areas of land-use types and develop training sites (Thapa et al. 2014).
The training sites or signature is employed to verify and define distinct classes (Jenson 1996).This is achieved either by user's prior knowledge of the geographic features of an area of interest such as identification of distinct, homogenous regions that represent each class (e.g., water or grass) or by ground-truth data such as GPS points, which refers to the acquisition of knowledge about the study area from field work, analysis of aerial photography, and from personal experience (Conner et al. 1975).Ground-truth data are considered to be the most accurate (true) data available about the area of study.They should be collected at the same time as the remotely sensed data, so the data corresponds as much as possible (Stars & Estes 1990).Furthermore, elements of visual interpretations such as color, shape, texture and pattern on aerial photos are commonly used that provides valuable clue during supervised classification.For example, we employed a texture and pattern analysis technique on aerial photos and selected pixels in such areas.With texture and pattern it is easy to differentiate naturally growing trees and human managed plantations, e.g., coconut and pine plantations.We derived seven LU/LC categories (water, urban areas, pine forest, pine plantation, mixed forest, grass, and cypress forest on floodplain) from the Landsat image (Fig. 2).We classified entire pixels into their designated classes according to the vegetation categories found in the study area.For example areas with tall pine trees were classified as 'pine trees', areas with mixed pine and oak trees were labeled as 'mixed forest', areas of floodplain were labeled as 'cypress trees on sloughs' and so on.GPS locations of each category accompanied with aerial and field photos were extensively used during classification process.

Accuracy Assessment
It is necessary to assess the accuracy of any thematic classification to evaluate its intended application, and high accuracy assures consistency and reliability of derived landscape metrics (Xulong et al. 2005;Shao & Wu 2008).Several factors related to the sensors as well as to the classification process contribute to classification errors (Lunetta et al. 1991).It is also critical to measure the quality and accuracy of data used for classification (Congalton & Green 1999).The classification or errors are analyzed by a confusion or error matrix, which is also called accuracy assessment (Congalton & Green 1999).An error matrix or accuracy assessment cell array is a table with entries representing the number of sample units; i.e., pixels, clusters of pixels, or polygons assigned to a specified class relative to the actual class found on the ground (Congalton 1991).Rows contain a list of class values for the pixels in the classified image file and columns represent class values for the corresponding reference pixels, determined by input from the user collected from sources such as aerial photographs, GPS points, previously tested maps or other data.The reference class values are compared with the classified image class values to assess the accuracy of an image classification.According to Anderson et al. (1976), classification accuracy close to 85% is acceptable for a LU/LC study.
Several statistical measures of a classified LU/LC map can be derived from an error matrix, including overall classification accuracy (sum of the diagonal elements divided by the total number of sample points), categorical omission and commission errors, and the KHAT coefficient (an index that measures the agreement between reference and classified data i.e., KHAT=1 when the agreement between reference and classified data reach 100%).A minimum of 204 reference points are required to achieve 85% accuracy with an allowable error of 5% (Jensen 1996).First we generated about 300 random (reference) points, and with help of aerial photos and with prior knowledge of geographic features we assigned values for each random point.Then we compared these reference class values with the classified image class values, which gave us an overall accuracy of 77.33% with KHAT = 0.7277.Then we used GPS locations as reference points and compared them with the class values of image files, which produced an overall accuracy of 81.48% with KHAT = 0.7449 (Table 1).The latter accuracy was deemed acceptable for this study because it was within the 5% allowable margin of error and was closer to 85%.

Habitat Suitability Index Models
We computed the HSI value for each pixel of the resultant classified image according to the following procedure.We selected the pine trees class because this habitat is required for successful breeding and foraging.We assigned a value of 1 to this class and set all others to 0, producing a binary map showing pine trees only.We ran neighborhood analysis in ArcGIS, which is a statistic configurations and compositions of a landscape (Doster et al. 1998;Thapa et al. 2014).

Image Processing
From a total of 287 GPS points, we used 162 points for classification accuracy.The rest (125) were used to classify the image.Use of GPS points and visual interpretation of aerial photos proved effective in Landsat ETM+ classification and facilitated the process.Our results show pine trees and grass have the lowest classification accuracy with 74.58% and 75% respectively.For pine trees this might be due to insufficient GPS points, because we could not gather data from private land containing pure stands of old pine trees.For grass it is possible to include agricultural lands, a common problem with landsat images having 30m resolution.In addition, short-grass areas (grazed pastures or manicured lawns) and dirt roads had overlapping values with other urban areas such as patches of bare soil and asphalt roads.We classified pine plantations with 90.91% accuracy because they were easily identified based on texture and pattern on aerial photos and GPS data collected from within plantation areas.Similarly, we classified urban areas with 89.66% accuracy as they are also easily identifiable on aerial photos and GPS data.Water pixels were classified with 100% accuracy.And it is one of the geographic features that a user can accurately classify in remote sensing applications as water pixels exhibits the lowest reflectance property when examined in a spectral profile.Profile Tools of ERDAS allow the users to examine spectral behavior of pixels of different features.Cypress trees on sloughs class was classified with 76.92% accuracy.Cypress trees occur mostly in sloughs of floodplains mixed with oak species and several factors such as topographic shadows and deep water contributed to the low accuracy of this class.Similar to water, wet sandy banks of creeks, streams, rivers, and sloughs have a lower reflectance in most bands and they became a source of confusion.Statistical analysis of spectral responses or profile from training samples, as well as ellipse and dendrogram plots, showed a similar reflectance with dark pixels (wet soils, black soil, and topographic shadows).Furthermore, vegetated (forest, urban, grass) and non-vegetated (water) were spectrally distinct.In order to redefine, refine and improve accuracy, we constantly reduced, merged and masked the confused classes.

Metrics at Patch, Class, and Landscape Levels
We calculated two metrics i.e. area/density/edge and connectivity of three different patch types (class) or habitat types: very unsuitable, unsuitable and potentially suitable (Table 2).These metrics were used to examine composition and configuration of patches in the study area (McGarigal et al. 2002).FRAGSTATS provides individual patch properties at three levels: patch, class and landscape, but we quantified patch properties at class level only because most metrics are redundant and provide similar values at patch and landscape levels.For example, total core area (TCA) at of total landscape occupied by the largest patch.LPI = 0 when the largest patch of the corresponding patch type is small and 100 when the largest patch occupies the entire landscape.The largest patch of potentially suitable habitat occupies only 0.1-0.42% of the landscape as compared to 30-60 % and 18-40 % for very unsuitable and unsuitable habitat types respectively.This corroborates the results of CA/TA and PLAND that showed presence of small amount of potentially habitat types as compared to the other two.
Connectivity is considered a vital element of landscape structure, and we used a single connectivity metric, COHESION, to observe physical connection between patches.COHESION = 0 when patches are less connected and approaches 100 when they are more connected.Our analysis showed the potential suitable habitat patch type is physically disconnected as indicated by 93-96 as compared to the other two habitat types with sometimes approaching almost 100 showing they are more connected and contiguous.

Image Processing
Overall, the LU/LC map derived from satellite imagery was satisfactory because categories were adequately mapped and resulted only in minor misclassifications.
The resultant map was refined with spatial masking and recoding to achieve acceptable accuracy.Use of aerial photographs and GPS points proved effective in improving classification accuracy.The contrasting reflectance of bare areas and vegetation in the visible and infrared bands facilitated accurate identification.However, accurate delineation of grass from crops and shrubs represented a challenge (as in many remote sensing studies).Visual examination of the satellite imagery of the study area and field work revealed numerous dirt roads crisscrossing the entire landscape.Our study area once contained booming oil towns and clearly shows signs of human-induced fragmentation.Several pipelines, power lines and railroads cut through the study area, dissecting the landscape into smaller fragments.
GPS locations of different categories or classes proved to be the most critical data during LU/LC classification of the landsat image in facilitating and enhancing