Mapping Regional Forest Management Units: A Road-based Framework in Southeastern Coastal Plain and Piedmont

Management practices are one of the most important factors affecting forest structure and function. Landowners in southern United States manage forests using appropriately sized areas, to meet management objectives that include economic return, sustainability, and esthetic enjoyment. Road networks spatially designate the socio-environmental elements for the forests, which represented and aggregated as forest management units. Road networks are widely used for managing forests by setting logging roads and firebreaks. We propose that common types of forest management are practiced in road-delineated units that can be determined by remote sensing satellite imagery coupled with crowd-sourced road network datasets. Satellite sensors do not always capture road-caused canopy openings, so it is difficult to delineate ecologically relevant units based only on satellite data. By integrating citizen-based road networks with the National Land Cover Database, we mapped road-delineated management units across the regional landscape and analyzed the size frequency distribution of management units. We found the road-delineated units smaller than 0.5 Ha comprised 64% of the number of units, but only 0.98% of the total forest area. We also applied a statistical similarity test (Warren’s I) to access the equivalency of road-delineated units with forest disturbances by simulating a serious of neutral landscapes. The outputs showed that the whole southeastern U.S. has the probability of road-delineated unit of 0.44 and production forests overlapped significantly with disturbance areas with an average probability of 0.50. does use affect forest


Introduction
The Southeastern United States (SEUS) forest comprises 32% of the total U.S. forestland (Oswalt et al, 2014), which combined with the productivity of the forest, places this region at the forefront of American forestry production (Fox et al, 2007). This heterogeneous landscape is composed of heavily managed forests, intensive agriculture, and multiple metropolitan areas.
SEUS, although one of the most densely forested regions in the United States (Hanson, 2010), is also heavily dissected by road networks (Coffin, 2007). The diverse forest management patterns, reflecting long-term land-use legacies (Haynes et al, 2002;Josephson et al, 1989), contribute to the complex land mosaic of SEUS.
It is challenging to quantify the ecological and anthropogenic mechanisms that control the spatial structure of the forest landscape and its surrounding areas in these complex forest mosaics. Forest management is the predominant factor in forest ecology and structural patterns (Becknell et al, 2015), but little is known about how management practices are related to surrounding land-use at the regional scale. One thing that is known is that, in the SEUS, significant expansions of urban areas tend to convert forested land to urban uses and that croplands tend to transition to pine plantations (Becknell et al, 2015, Haynes, 2002Greis, 2002 andStanturf et al, 2003). To understand the ecological and anthropogenic influences of differently managed forests on ecosystem processes, all landscapes should be understood at multiple scales, from the local scale (forest management unit) to regional scales, with the regional scale referring to broad forest mosaics that are formed from management patches (O'Neill et al, 1996). Understanding forest management spatial patterns require defining a map-based management unit, which is the subdivision regarding effects of land use on forest ecosystems. This research seeks to further the understanding of how spatial patterns of forest be classified into four categories: production forestry, ecological forestry, passive management and preservation management (Becknell et al 2015). Production management harvests forest products and sustains the bio-productivity of the system with the sole objective of producing wood, pulp, and other forest products. In SEUS, production management based on silviculture systems, which homogenize parts of the landscape, has predominated the SEUS (Siry et al, 2002). However, forest conservation systems have evolved considerably over recent decades (Mitchener et al, 2005;Franklin et al, 2007). Ecological management uses legacies of disturbance, including intermediate stand disturbance processes such as variable density thinning and fire, and variable and appropriate recovery times to manage forests that still produce economically valuable wood products while preserving many of the values of natural forests (Franklin et al. 2007). Passive management is defined as a practice with little or no active management. We argue that all forests are managed to some degree and that doing nothing is a form of management. Preservation forestry aims to minimize the ecological footprint of society with the objectives of protecting wildlife and maintaining ecosystem services. Furthermore, certain forest management practices can suppress wildfires (Waldrop et al, 1992), prevent insect /pathogen outbreaks (Netherer et al, 2005;Faccoli et al, 2014), change water yield and hydrologic regulation (Douglass, 1983), produce wood products, provide places for hunting and recreation, and conserve habitat biodiversity.
Forestland structures, functions, and ecological processes are scale-dependent (Niemelä, 1999;Drever et al, 2006;Battaglia et al, 1998 Road networks facilitate movements of humans and connect natural resources with societies and economies. As conduits for human access to nature, the physical footprint of approximately 6.6 million km of roads in the United States (U.S. Department of Transportation. 2013) has significant primary and secondary impacts on ecosystems and the distribution of species (Bennett 1991). Fifteen to twenty percent of American land is subject to the ecological effects of road networks Alexander 1998, Forman et al. 2000). The most noticeable effects of road networks on forest structures are landscape structure changes, including reduced mean patch size, increased patch shape complexity, increased edge densities, and reduced unit connectivity. In one case, McGarigal et al. (2001) investigated the landscape structure changes of the San Juan Mountains from 1950-1993 and found that roads had a more significant ecological impact (e.g. core forest areas and patches sizes decrease) on landscape structure than logging activities.
In addition to management practices (e.g., harvesting, fertilizing), the construction of road networks divides the forested land into smaller patches, thereby increasing the potential intensity of the effects of management practices. Road networks in managed forests provide easy access for managers and harvesters to extract and regenerate resources (Demir et al, 2005).
Roads may influence fire regimes by increasing fire ignition as a result of human activities (Franklin and Forman, 1987). Moreover, road networks alter the spatial configuration of management patches by functioning as firebreaks, which form new patterns in landscapes (Franklin and Forman. 1987;Nelson and Finn, 1991;Eker and Coban, 2010). By quantifying the spatial patterns of management units created by roads we may gain insight into the ecological effects of road networks on spatial forest structures within differently managed areas.
The impacts and ecological effects of roads on the landscape might be misestimated because methods measuring the road-effect zones and landscape scale effects are not yet well developed (Ries et al, 2004, Hou et al, 2013. Roads and streams may be challenging to identify, or invisible because they do not open the canopy so that many or roads are not detectable on satellite imagery or even aerial photography. The reliability of large-scale road data is also challenged due to issues of accuracy, coverage and immediacy, all of which can underestimate the extent and ecological impacts of roads on forest structures (Riitters et al, 2004). We propose that common types of forest management are practiced in road-delineated units that are detectable by remote sensing satellite images coupled with crowd-sourced road network datasets.
We also describe and study the patterns of forest management units in response to land ownership and different management practices.

Study Area
We focus on the Southeastern U.S. Coastal Plain and Piedmont (SEUS) region ( Figure 1).
The SEUS is located between Piedmont to the north and the Atlantic Ocean to the east and covers a significant portion of the southeastern United States. The SEUS is the home to the most densely production-forested region nationwide, which makes up 32% of total U.S forest cover (Oswalt et al. 2014). Based on EPA eco-region descriptions, land cover in the SEUS is a mosaic of cropland, pasture, woodland and forests (Bailey, 2004). Major silvicultural forests in SEUS are pine forests, such as slash pine (Pinus elliottii Engelm.) and loblolly pine (Pinus taeda L.) forests. European settlement and the extensive harvesting in the early 1900s removed 98% of the original longleaf pine (Pinus palustris Mill.) forests, which was one of the most dominant ecosystems in SEUS (Outcalt, 2000) and converted them to plantations of native slash pine. The SEUS forest system is a fire-dominated system with native trees adapted to short-period standclearing events. The primary forest management types are production and passive management due to the dominant ownership of private owners, logging companies, and investment institutions (Real Estate Investment Trusts (REITs) and Timber Investment Management Organizations (TIMOs)) (Zhang et al, 2012). This diversity of land cover types is spatially heterogenous, and patch sizes of the numerous vegetation classes vary across a wide range of scales ( Figure 1).

Forest Extent
In this study, forest extent is determined by a composite of the 2006 and 2011 USGS National Land Cover Database (NLCD), which was constructed from Landsat imagery at 30-m spatial resolution (Jin et al, 2013). We aggregated 21 NLCD classes into two classes: forest (deciduous, evergreen, mixed forest and woody wetland) and non-forest (including water). Only the pixels that contain 50% or more forest area in NLCD will be considered as forested pixels.
We also extracted the SEUS urban areas by using the most recent 2015 US Census Bureau's TIGER cartographic boundary urban areas (TIGER, 2015) dataset to remove urban areas from theanalysis.

Forest Management Type
A map of forest management type was produced in earlier work (Marsik et al, 2018). An integrated random forest classifier was built from the analysis of long-term phenological features derived from BFAST outputs and spectral entropy calculated from the Terra-MODIS enhanced vegetation index (EVI: MOD13Q data product), along with ancillary data such as land ownership, and disturbance history to classify different forest management types (Breiman, 2001;

Road Networks
We selected OpenStreetMap as the primary road data and the USDA National Forest service trail and road maps Road density in SEUS was measured as the total length of all roads (in kilometers) in a district divided by the total land coverage area of the district (km 2 ) based on our developed road networks map ( Figure S2).

Landscape Fire and Resource Management Planning Tools (LANDFIRE)
Disturbance data from the Landscape Fire and Resource Management Planning Tools (LANDFIRE) disturbance database were used to evaluate the management unit map.
LANDFIRE is a combination of Landsat images, fire program data, and cooperator-provided field data and other ancillary databases (e.g., PAD-US)., and is a shared program between the wildland fire management programs of the U.S. Department of Agriculture Forest Services and U.S. Department of the Interior (Rollins, 2009). LANDFIRE also describes land cover/use both spatially and temporally from 1999 to 2014 and provides the existing vegetation composition map based on dominant species or group of dominant species. Spatially, LANDFIRE is a Landsat-based (30 m) database, which matches the 30-m spatial resolution of this study (https://www.landfire.gov/disturbance.php). We chose the LANDFIRE project data because the spatial scale is small enough to detect subtle changes brought about by land management practices, and large enough to reflect the characteristic variability of essential ecological processes (such as wildfire) in the appropriate spatial context. The disturbance data from LANDFIRE will be used to evaluate management units delineated by the road network in the context of intensely managed forests in SEUS (see Section 4.2, Figure S3).

Forest Ownership
Geospatial land-ownership data from federal and nongovernmental agencies were integrated for land ownership mapping ( Figure S4). Forest ownership in SEUS is broadly categorized as publicly owned and privately owned according to the landowners. There are six sub-types of public ownership, which are federally protected, federal, state protected, state, military, and local. Also, there are four sub-types of private ownership: non-governmental organization, private, family, and corporate. The ownership classification implies different management objectives, as well as landowner skills, budgets and interests. Datasets from other federal and state government agencies were regrouped and classified into ten sub-types to create a comprehensive dataset that includes public land ownership and privately protected easements as well as specially designated areas and associated protection level (see Table S1). The final product is a 250-meter spatial resolution raster data depicting the forest ownership types and resampled to 30-m in this study to match the spatial resolution of the NLCD database.

Mapping Road-delineated Units
In each forest management type, the fundamental element of management practice is the management unit. In this study, we define the individual forest clusters that delineated by roadnetworks as "management units" and the clusters that directly derived from forest extent map as "management patches." We hereby developed two comparative methods for landscape analysis by using two sets of input data, with and without incorporating OpenStreetMap. For the method without incorporating OpenStreetMap, management patches were mapped on the forest extent map resulting from the map described in section 3.1 and the Region Group tool in ESRI ArcGIS 10.X (ESRI Inc,) to identify clusters of forest pixels that formed unique and unconnected forests.
Road-delineated forests units were mapped on the forest extent map from section 3.1 after superimposing detailed road networks with the Region Group tool to identify forest clusters as units. When superimposing road maps, all road networks were converted to one-pixel segments. After one-pixel wide road segments were derived, we converted all the forest pixels to 1 and the pixels that contained at least one road segment to non-forest pixels (30-m spatial resolution) to 0.

Geospatial Assessment
In SEUS, regional forest management activities were represented as disturbances as described by LANDFIRE data, such as clear cuts, fires, and thinning. Figure 2 shows the example view of LANDFIRE cumulative disturbance with delineated forest extent, and it can be clearly observed (with the Warren's I of 0.62) that the disturbed areas and delineated unit shared boundaries and show a large degree of equivalence. The spatial coincidence has been shown to facilitate the interpretation and integration of defining regional forest management units. In this study, we propose a test based on forest management unit and the geographic corresponding forest disturbances to compare the geographical similarity between the management unit and corresponding forest disturbances. The assessment of geographic image overlap is analogous to quantifying the niche similarity of two species in two dimensions. As two-dimensional rasters, both disturbance and forest extent data can be treated as homogenous and spatial-explicit datasets.
Testing the overlap between pairs of road-delineated forest extent with disturbance regions were compared using the similarity statistics of Warren's I (Warren et al, 2008). The (2) We further tested the hypothesis of road delineated forest management units by generating neutral landscape models from two perspective as: 1) Spatially. In this study, we used the Worldwide Reference System 2 (WRS-2) row 17 path 39 (17_39 for later), which is one scene of Landsat Thematic Mapper (TM). As shown in Figure   3, the forests in 17_39 are heavily disturbed and fragmented, but it also shared the larges 2) Iteratively. We applied a range-based attribute filter to the SEUS road density map ( Figure   S2). The road densities range from 0 to 91.48 km/km 2 , so we used the interval of 5 to randomly select a serious spatial noncorrelated 10×10 km grids with a total number of 19 landscapes in SEUS ( Figure 4). For each spot, we iteratively run the Random NLM algorithm for 500 times and calculated the Warran's I with overlaying the simulated disturbance resulted from NLM with the road networks delineated forest extent map.

Spatial Assessment
We calculated the Warren's I with overlaying the road networks delineated forest extent with LANDFIRE disturbance map on a 10×10 km grid in SEUS ( Figure 5). The total number of 10×10 km grids is 11,072. Figure 5 shows region after replacing the LANDFIRE data with simulated disturbance layer, as shown in Figure   6(d). The mean value of the probability distribution is 0.14 with the standard deviation of 0.082.
The results show that human-derived forest disturbances can result in a non-random association with forest extent. By comparing with the LANDFIRE disturbance derived Warrans' I map Figure 6(a), road networks make a great contribution in identifying and shaping forest patterns in the case study area 17_39, wherein SEUS reflect the management practices on the ground, and so for the road-delineated forest compartments, we hereby call them forest management units.
Road density in SEUS ranges from 0 to 91km/km 2 . Table 1  shows the evidence that roads help to shape forest patterns by statistically testing and approve the hypnosis that road delineates forest management patterns. The SEUS forests is dominated by management unit size ranging from 100 to 10,000 Ha.

Forest Management Unit map
The forest management unit-size class map was reclassified based on the forest management unit sizes ( Figure 8). From Figure 8, it can be clearly seen that riparian forests stand out as large, unbroken linear features (most of the orange colored units), which cover average unit sizes from 10,000 Ha to 100,000 Ha. A characteristic feature for the southeastern forest is that relatively small and large management units locate close to each other, surrounded with small-sized units ( Figure 8 and 10). Figure 10 shows To illustrate the contributions of road networks to our regional SEUS management unit map, a representative comparison set was done with and without incorporating road networks data. In Figure 11, we overlay the histogram of the management unit (incorporating road networks) with the histogram of the management patches (without incorporating road networks.
The red histogram shows the size-based frequency distribution of patches without incorporating road networks and the purple histogram illustrates the size frequency of forest management units.
When refining the management units with road networks, there are 17 times more patches (management patches) compared with the map without incorporating road networks (management units)

Management Units Under Different Management Approaches
Road density ( Figure S2) has been proposed as a broad index of roads' ecological effects in a landscape (Forman et al, 1998).  (Table 2).
We incorporated Warren's Index to assess quantitatively the geographic overlap between forest management units under different functional forest management types. For ecological management, the specific practice is designed to emulate the outcome of natural disturbance, which is to create an uneven-age stand structure to manage competition between and within multi-cohort stands. The distribution of ecological management units shows spatial heterogeneity with structurally complex stands. For passive management forest lands, as the passively managed forests mostly adopt many irregular shapes with blurred boundaries, and rupture of connectivity. For preservation management forest: mostly large government-managed land for multiple-purpose including watershed, wildlife, recreation and wilderness aspects. Accordingly, various practices may be applied to it such as harvest, cutting, retention cutting, thinning and prescribed fire. As shown in Table. 3 Warren's I represent the overlap between forest management units under different management types with the corresponding forest disturbance area.
The 10×10 km based spatial grid analysis of Warren's I is shown in Figure 5 and Figure (Haines et al, 2001). In this study, we used a threshold of 50% of the similarity score although many useful criteria for establishing such thresholds have been proposed (Jimenez-Valverde and Lobo, 2007). In Table 3's analysis, the criteria were just used to show binary predictions of four differently managed forests.

Ownership Representation of Forest Management Unit
Forest management activities are important links between human and environmental factors, especially at regional scale. Forest ownership patterns also explain different types of land management practices and trajectories of land cover change (Turner 1996). The aim of this part of the research is to link regional land ownership to management. We produced the SEUS ownership database ( Figure S4; Table S1) by collecting the data from different sources, where the ownership was divided as Private and Public. Based on our understanding of the SEUS forest ownership, we reclassified the forest owner types into public and private with 10 sub-classes ( boundaries, so there should be a spatial coincidence between road-delineated management units and disturbance. In this study, we assessed the forest management from stands to regional scale, by incorporating road networks and multi-temporal disturbance remote sensing database.
A number of conclusions can be drawn from the analysis in this study: 1) Road networks play a role in delineating forests from local to regional scale. By defining the individual forest clusters delineated by road networks as "forest management units" and the clusters that directly derived from forest extent map as "management patches", we mapped the forest extent map of both "units" and "patches" and compared them with treating "patches" map as background. There were 17 times more "units" than "patches" over the whole SEUS. And we also summarized the size distribution road delineated units, with units smaller than 0.5 Ha comprised 64% of the counts of units, these small units altogether covered only 0.98% of the total forest area.
2) We quantitatively tested the probability distribution patterns by using Warren's I of roaddelineated management units and the corresponding forest disturbances area. The average probability of road-delineated management units is 0.44, and we also visualized the probabilities by setting a 10km×10km grid. In SEUS, the high equivalency between the roaddelineated units and the corresponding areas were found at most production forests, and large-size preserved areas (e.g. Okefenokee National Wildlife Preserve and St. Marks National Wildlife Refuge).
3) The combination of remote sensing data and OpenStreetMap constitutes a useful tool to monitor, characterize and quantify land cover and management unit distributions at macrosystems scale. By using the NLCD as the forest reference data and OpenStreetMap as road networks dataset, we produced the OpenStreetMap refined management unit pattern map and analyzed the spatial size distribution of forest patterns. In addition, by incorporating

Availability of data and materials
The dataset and code used during the current study are available at: https://doi.org/10.6084/m9.figshare.11406612.v1

Consent for publication
Not applicable

Completing interests
The authors declare that they have no competing interests     Figure 3. Case study region of Worldwide Reference System 2 (WRSII) row 17 path 39 (17_39). As the LANDFIRE EVT product shows, it encompasses a diversity of landcover and disturbance types. Much of the conifer and conifer-hardwood land cover found outside of the large riparian area of the Okefenokee Swamp (upper central) are heavily managed, privately owned plantations and mixed agriculture/timber land use.      Figure 11. Comparison between with and without incorporating OpenStreetMap on unit size frequency distribution.