Mapping land cover and estimating...
Mapping land cover and estimating forest structure using satellite imagery and coarse resolution lidar in the Virgin Islands Todd A. Kennaway,a Eileen H. Helmer,b Michael A. Lefsky,c Thomas A. Brandeis,d Kirk R. Sherrillc aDepartment of Forest, Rangeland and Watershed Stewardship, Colorado State University, Fort Collins, CO 80523-1472, USA email@example.com bInternational Institute of Tropical Forestry, USDA Forest Service, Jard��n Bot��nico Sur, 1201 Calle Ceiba, R��o Piedras, PR 00926-1119, USA cCenter for Ecological Analysis of Lidar, Department of Natural Resources, Colorado State University, 131 Forestry Building, Fort Collins, CO 80523-1472, USA dSouthern Research Station, USDA Forest Service, 4700 Old Kingston Pike, Knoxville, TN 37919-5206 Abstract Current information on land cover, forest type and forest structure for the Virgin Islands is critical to land managers and researchers for accurate forest inventory and ecological monitoring. In this study, we use cloud free image mosaics of panchromatic sharpened Landsat ETM+ images and decision tree classification software to map land cover and forest type for the Virgin Islands, illustrating a low cost, repeatable mapping approach. Also, we test if coarse-resolution discrete lidar data that are often collected in conjunction with digital orthophotos are useful for mapping forest structural attributes. This approach addresses the factors that affect vegetation distribution and structure by testing if environmental variables can improve regression models of forest height and biomass derived from lidar data. The overall accuracy of the 29 forest and non-forest classes is 72%, while most the forest types are classified with greater than 70% accuracy. Due to the large point spacing of this lidar dataset, it is most appropriate for height measurements of dominant and co-dominant trees (R2 = 72%) due to its inability to accurately represent forest understory. Above ground biomass per hectare is estimated by its direct relationship with plot canopy height (R2 = 0.72%). Keywords: Land cover, decision tree software, discrete lidar, forest structure, regression modeling, Virgin Islands. 1 INTRODUCTION Information on land cover, forest type and forest structure for the Virgin Islands is limited to maps of ecological zones and photo interpreted land cover [1,2] and relatively few forest inventory plots. The lack of current land cover data and robust techniques for updating that data, and the sparseness of forest inventory plots relative to the number of different forest types, pose challenges to land managers and researchers in ecologically vulnerable subtropical environments. These challenges are made more acute in Caribbean environments, because the interaction between trade winds and steep topographic gradients cause forest types to change over short distances  and high rates of disturbance lead to variable forest structure. In this study, datasets of land cover and forest type are derived from satellite imagery with decision tree software, illustrating a low cost, repeatable approach for creating Journal of Applied Remote Sensing, Vol. 2, 023551 (12 December 2008) �� 2008 Society of Photo-Optical Instrumentation Engineers [DOI: 10.1117/1.3063939] Received 13 Jul 2008 accepted 3 Dec 2008 published 12 Dec 2008 [CCC: 19313195/2008/$25.00] Journal of Applied Remote Sensing, Vol. 2, 023551 (2008) Page 1
such data. Although decision tree classification is becoming common in remote sensing, only a few studies use decision trees for detailed forest mapping of subtropical islands [4-6]. This study also addresses the characterization of forest structure with airborne light detection and ranging (lidar) when inventory data are sparse. Lidar adds a third (z) dimension to the spatial description of forest types with accurate estimates of vegetation height and above ground biomass . No prior research has addressed the quantification of lidar derived forest structure in the Virgin Islands. Data describing indices of forest structure such as height and biomass can provide important information such as indicators of forest age, species richness and habitat. Forest distribution and structure in the Virgin Islands has been modified for hundreds of years by both natural and human caused disturbances, including hurricanes and human exploitation. Prior to European colonization, indigenous peoples such as the Carib first cleared forest for food, shelter, and boat building materials. In the 1600 and 1700���s Danish and British settlers arrived and began converting forest to intensive agriculture that included coffee, sugar cane and tobacco [8,9,10,11]. As a result, most old growth forest was cleared and has recovered as fragmented secondary forests after the gradual abandonment of agriculture through to the early twentieth century. The forest clearing had lasting impacts on forest structure, ecosystem function and species composition, including the introduction and extinction of exotic and endemic species . Increased pressure from urban development has led to additional forest clearing in the Lesser Antilles  and Puerto Rico. The spatial pattern of forest clearing is often influenced by proximity to existing urban areas, roads and topography [13,4]. Islands such as St. Thomas  and Tortola have also experienced urban growth at the expense of forested areas over the last decade. About 65% of St. John is protected by the US Park Service including much of its semi-deciduous (including semi-evergreen) and deciduous forests. However, the unprotected low elevation dry forests on that and other islands, which have been shown to be important habitat for many avian species, are considered endangered and susceptible to increasing developmental pressures . The overall goal of this study is to develop an approach for characterizing the structure of varied subtropical island forest formations when available inventory data are relatively sparse. To accomplish this goal, we developed three main objectives. The first objective is to test an improvement to a previously developed approach for using Landsat image mosaics to map land-cover and forest types in persistently cloudy, complex tropical landscapes with decision tree classification software [4-6,16,17]. The improvement is that we use panchromatic- sharpened image mosaics to increase spatial resolution in the resulting maps. We also test whether the approach is applicable to a large area, the Virgin Islands, which includes many islands. The second objective is to test if the coarse-resolution (shot spacing of 2.76m) discrete lidar data that are often collected in conjunction with digital orthophotos are useful for mapping forest structural attributes, including height and biomass, over the steep environmental gradients present on the islands of St. John and St. Thomas. Also, this study tests whether integrating Landsat ETM+ satellite imagery and environmental variables with the lidar data can improve models of forest structural attributes. Whether the large point spacing of such coarse resolution lidar will be adequate to accurately sample and model forest structure parameters has not been tested. Also unknown is whether the range of physiognomic types found across these islands will complicate the estimates of forest height and biomass. Several studies have shown that large footprint scanning lidar accurately predicts forest structure, including canopy height, basal area and above ground biomass in Journal of Applied Remote Sensing, Vol. 2, 023551 (2008) Page 2
Douglas fir/western hemlock forests [18,19]. Other studies accurately model forest metrics with discrete lidar, focusing on small tracts of homogeneous forest stands and small foot-print sensors . Accurate estimates of forest structural attributes using regressions have been successfully performed such as height [21,22], aboveground biomass [23,24], and crown diameter [25,26]. Lidar based biomass has been estimated for a variety of forest types including, but not limited to, temperate mixed deciduous coniferous [25,23,27], temperate deciduous  and tropical rain forests . The third objective is to summarize forest structure of the predicted forest structural attributes for each mapped forest class. This step allows us to characterize forest height and biomass for different forest types on St. John and St. Thomas. The datasets generated in the project will support other studies in the Virgin Islands, including avian monitoring surveys and the Forest Stewardship Program . 2 METHODS 2.1 Study area The US and British Virgin islands (18��20���N, 64��40���W) are a part of the Caribbean���s Lesser Antilles and are composed of six major and 40+ minor islands and cays. The major islands in the US territory include St. Thomas, St. John and St Croix, while the main islands in the British territory include Tortola, Virgin Gorda and Anegada (Figure 1). The islands have a combined area of about 50,000 ha, with subdued to rugged topography and elevations ranging from just below sea level in some wetlands to over 500 m on the island of Tortola. The climate is mostly subtropical, with a hot and humid rainy season that extends from May to November and a dry season that is tempered by trade winds. The geology of the islands consists of alluvial, sedimentary, volcanic and limestone strata. Ecological zones on the islands include Subtropical Moist and Dry forest sensu Holdridge [1,30]. Lidar Study Area St. John and St. Thomas were selected as the lidar study area based on the availability of lidar data coverage. The island of St. John (18��22���N, 64��40���W) and the island of St. Thomas (18��21���N, 64��55���) are about 5,000 and 7,200 hectares in area, respectively, and consist of mountainous topography with elevations ranging from sea level to 387 m on St. John and 471 m on St. Thomas. The woody vegetation on both islands is similar to other islands in the Virgin Islands and includes both late and early stage successional forests. In 1956, the US Park Service established the Virgin Islands National Park (VINP). Protecting about 65% of St. John, it includes the island���s interior high elevation semi- evergreen and deciduous forests. The long standing reserve status has helped protect most of the island���s forests from development, creating one of the largest contiguous expanses of forest in the Lesser Antilles. The VINP provides unique research opportunities to study the island���s diverse ecology and establishes a template for monitoring mature successional forest structure. In contrast, the forests of St. Thomas which make up about 69% of the island has not received protection status and developmental pressures and impacts can be observed island-wide. 2.2 Landsat Imagery and Reference Data A land-cover and forest type map for the US and British Virgin Islands was created by supervised classification of Landsat ETM+ imagery using decision tree analysis software. An image mosaic for about the year 2000 was created from Landsat scenes of various dates. The Journal of Applied Remote Sensing, Vol. 2, 023551 (2008) Page 3
Ginger Island Tortola St Thomas St John St Croix Anegada Virgin Gorda Jost Van Dyke Beef Island Peter Island Guana Island Great Camanoe Norman Island Hans Lollik Island Cooper Island Thatch Island Great Tobago Savana Island Saba Island Buck Island 64��30'0"W 64��30'0"W 65��0'0"W 65��0'0"W 18��24'0"N 18��24'0"N 18��0'0"N 18��0'0"N C a r i b b e a n S e a 10 0 10 20 5 Kilometers US and British Virgin Islands Area in Detail C a r i b b e a n S e a A t l a n t i c O c e a n Elevation (m) High : 531 Low : -18 Low : 0 N Fig. 1. Map of the study area. reference scenes for the mosaic were World Reference System 2 Path/Row 004/047-048, both dated 27 Mar 00. The scenes used to fill cloud-masked or edge areas in Path/Row 004/047 were dated 02 Nov 01 (Path/Row 003/047), 17 Sep 99, 02 Aug 00, and 25 Jan 01. The scene used to fill cloud-masked areas in Path/Row 004/048, was dated 25 Jan 01. Cloud obstruction in the reference image was 20.9 % before and 5.3% after the cloud removal and mosaic process. The 30-m multispectral bands for each scene were first cloud-masked and then matched to the reference scene with regression tree normalization . This technique models the relationship between co-located pixels from different image dates and estimates new image digital numbers (DNs) to fill in the cloud and cloud-shadow masked areas of the reference scene. In addition, the technique reduces atmospheric and phenological differences that occur with multi-date image mosaics . Likewise, the 15-m panchromatic band for each scene was also matched to the reference panchromatic band with regression tree normalization models based only on the panchromatic bands. The matched panchromatic Journal of Applied Remote Sensing, Vol. 2, 023551 (2008) Page 4
image parts were then mosaicked, and the panchromatic mosaic was then used to pan-sharpen the 30-m mosaic of the multispectral bands. Principal components transformation was chosen to merge the native 30 m Landsat multispectral image bands (bands1-5, 7) for each scene with the 15 m panchromatic band (band 8). It was chosen based on 1) results from preliminary tests of other resolution merging methodologies (Brovey and Multiplicative) available in ERDAS Imagine, and 2) other studies [32-34] that have concluded that the principle components transformation method provides increased spatial resolution without degrading spectral discrimination. Ancillary data were used to create an island-wide predictor variable dataset to assist in the classification of image pixels. Adding geographic data ancillary to satellite imagery improves classification of land cover and forest types by reducing spectral confusion among vegetation classes [35,36], including in Caribbean island landscapes . Topographic variables derived from United States Geological Survey (USGS) 30 m digital elevation models (DEM) for the US territory and 90 m Shuttle Radar Topography Mission (SRTM) elevation datasets resampled to 30 m for the British territory included elevation, slope and aspect [37,38]. Climatic variables included mean annual precipitation and temperature . Variables derived from USGS Digital Line Graphics (DLG) for the US islands and scanned topographic maps for the British islands at a scale of 1:24,000 that were registered to the image mosaic include distance to primary and secondary roads, distance to streams and ravines, and distance to coastlines . The ancillary predictor data was spatially co-registered with the cloud free image mosaic and stacked with the Landsat ETM+ reflectance bands 1-5, 7, and two band indices, resulting in an 18 band image mosaic for the classification. The band indices included the Landsat ETM+ image bands to produce the normalized difference vegetation index (NDVI) and 4/5 band ratio, which are useful indictors of vegetation vigor and forest structure [41-43]. Field surveys in 2005 and consultation with experts enabled us to discern land-cover and forest type in reference imagery and the classification image mosaic. The reference imagery included 1 m IKONOS panchromatic sharpened imagery for the US and British Islands and additional 1m color digital ortho quarter quads (DOQQ) for the US islands. Land cover and forest type were then identified in the satellite imagery. Forest type was identifiable in both the reference imagery and the Landsat imagery by color, tone and texture as well as spatial indicators including aspect and elevation. Difficulties distinguishing forest type were encountered in areas that were transitional between semi-deciduous and seasonal evergreen forest. Field survey data proved useful in identifying these transitional areas in the reference imagery. Training data for the image classification model was derived using the reference imagery and field data collected in the 2005 reconnaissance survey. About 25 to 250 multiple pixel polygons were distributed for each class throughout the extent of the study area. Training data samples collected over a large inter-island extent ensured thorough representation of each class and provided a full range of variability for the class. For example, there are often spectral variations in similar forest types where the image scenes were radiometrically matched in the cloud elimination procedure . Target classes included sunlit and shadowed woody vegetation types, sunlit and shadowed green and senescent pasture, mangrove, wetland, and non-forested classes (Table 1). We used the woody vegetation classification system designated at the formation level (Table 1) that  adapted for Landsat imagery classification from . Areas with less than Journal of Applied Remote Sensing, Vol. 2, 023551 (2008) Page 5