Extent and spatial patterns of grass bald land cover change (1948–2000), Oregon Coast Range, USA

Abstract

Globally, temperate grasslands and meadows have sharply declined in spatial extent. Loss and fragmentation of grasslands and meadows may impact biodiversity, carbon storage, energy balance, and climate change. In the Pacific Northwest region of North America, grasslands and meadows have declined in extent over the past century. Largely undocumented in this regional decline are the grass balds of the Oregon Coast Range, isolated grasslands in a landscape dominated by coniferous forests. This study was conducted to quantify the spatial extent and patterns of grass bald change. Five balds in the Oregon Coast Range were evaluated using historical aerial photographs and recent digital orthophoto quadrangles (DOQ). Over the time period of study (1948/1953 to 1994/2000), bald area declined by 66%, primarily from forest encroachment. The number and average size of bald vegetation patches declined, while edge density increased. Tree encroachment into balds was inversely related to distance from nearest potential parent trees. Spatial patterns of bald loss may result from a forest to bald gradient of unfavorable environmental conditions for tree establishment and/or seed dispersal limitation. Species dependent on balds may be at risk from loss of bald area and increased fragmentation, although metrics of habitat fragmentation may not reflect species-specific habitat requirements. Tree encroachment patterns and increased bald edge densities suggest increasing rates of bald loss in the future. The remote sensing nature of this study cannot determine the fundamental causes of bald decline, although prior research suggests climate change, cessation of native burning, successional changes in response to prior wildfires, and cessation of livestock grazing all may have potential influence.

Appendix A. Aerial photograph acquisition and image processing

Aerial photo acquisition

Historical panchromatic 1:12,000 scale aerial photographs were acquired from the University of Oregon Map Library in Eugene, OR. All historical aerial photographs were taken during the summer of 1948 (except Mount Hebo, taken during the summer of 1953). Historical aerial photographs were scanned on a flatbed scanner at 600 dpi. Scanned historical images had 8-bit radiometric resolution and 0.63–0.89 m postrectification spatial resolution, re-sampled to 1 m pixel size after geo-rectification using nearest neighbor interpolation. Recent images of the study areas are panchromatic digital orthophoto quadrangles (DOQs). DOQs were downloaded from TerraServer-USA (http://terraserver-usa.com/default.aspx). The Mount Hebo, Marys Peak, Grass Mountain, and Prairie Peak DOQs were taken in the summer of 1994 (except Bald Mountain, taken during the summer of 2000). DOQs were in Universal Transverse Mercator Projection (UTM, Zone 10N), using the Geodetic Reference System Spheroid of 1980, and the North American Datum of 1983. All DOQs had 8-bit radiometric resolution and 1 m spatial resolution.

Aerial photo geo-rectification

Each scanned historical image was first manually clipped so only grass balds and surrounding forest vegetation were retained. Each clipped image was geo-rectified using the recent DOQ of the same site as the reference image. Control Points (CPs) between historic and recent images were manually selected, focusing on the most temporally stable features (i.e., buildings, roads, rock outcrops, and individual open grown trees). A minimum of 20 CPs between historical and reference imagery was desired, although this was not possible in all cases. Geo-rectification models used polynomial equations, evaluated by the Root Mean Squared Error (RMSE), and rejected if the RMSE exceeded 0.5 pixels. Additionally, rectified images were digitally overlain as partial transparencies over recent DOQs to qualitatively assess rectification accuracy. Rectification and resampling occurred using the nearest neighbor method, conserving raw image brightness values. A mosaic of two rectified images was required for the Prairie Peak study area (Prairie Peak a & b), and was generated using histogram matching, no cutlines, minimum select function, and union of all inputs.

Study area	Date	CPs	RMSE X	RMSE Y	RMSE total	Polynomial order
Mount Hebo	1953	22	0.05	0.03	0.06	1st
Bald Mountain	1948	8	0.03	0.01	0.02	1st
Marys Peak	1948	11	0.02	0.29	0.04	2nd
Grass Mountain	1948	20	0.02	0.01	0.02	2nd
Prairie Peak—a	1948	24	0.06	0.06	0.08	1st
Prairie Peak—b	1948	20	0.04	0.02	0.05	1st

Image filtering, level slicing land cover classification, and matrixing

DOQs and rectified historical images were filtered using a 5 × 5 pixel focal minimum filter followed by a 5 × 5 low-pass filter. This filtering combination compensated for highly reflective (high brightness value) pixels within continuous forest canopy that would be misclassified during land cover classification as bald vegetation or bare ground unless their brightness values were reduced. The 5 × 5 focal minimum filter assigned the lowest brightness value of the neighboring 24 pixels to the central pixel. Focal filtering was followed by a 5 × 5 low-pass filter to deemphasize high-frequency edges (Lillesand et al. 2004).

Classification of pixels into three landcover classes (grass bald, forest, and bare ground) was accomplished by level slicing of gray-level brightness values. This method has previously been used on panchromatic images to classify pine forests and grasslands in the Colorado Front Range (Mast et al. 1997). Each filtered image was sliced individually, and brightness value ranges for each land cover class varied between each of the 10 images. Attempts to classify conifers and hardwoods separately were unsuccessful; instead a single forest land cover classification included both conifers and hardwoods. Using level slicing to classify the roads and buildings land cover class greatly reduced classification accuracy of the other three classes. Instead, the roads and buildings were manually delineated as polygons in separate shapefiles for each current and rectified historical image in ArcGIS 9.1 (2005 ERSI Inc.). Polygon shapefiles were converted to raster images and incorporated into each three-class classified image using an overlay function, resulting in four-class classified images.

Despite focal minimum and low-pass filtering, visual comparison of classified imagery to the geo-rectified images and DOQs found numerous instances where small areas of forest were misclassified as grass balds. Clumping and sieving was used to reclassify these false balds as forest. An eight pixel window was used to clump contiguous groups of bald pixels into individual bald patches. Bald patches were determined manually by overlaying the clumped bald patches over the associated rectified image or DOQ. Clumped bald patches that were smaller than the smallest visually determined patches within a historical or DOQ image were reclassified as forest, resulting in historical and recent 4-class (forest, meadow, bare ground, and roads and buildings) land cover images for each study area. Change detection of bald land cover 4 between historical and recent classified images used a matrix function, creating a new image showing the coincidence of values between the historical and recent classified images. Each matrix classified image had 16 potential coincident classes (4 classes × 4 classes). These classes were recoded, and the surface area of each land cover type was calculated from the number of pixels within each class.