Large-Area, High-Resolution Tree Cover Mapping with Multi-Temporal SPOT5 Imagery, New South Wales, Australia
Tree cover maps are used for many purposes, such as vegetation mapping, habitat connectivity and fragmentation studies. Small remnant patches of native vegetation are recognised as ecologically important, yet they are underestimated in remote sensing products derived from Landsat. High spatial resolution sensors are capable of mapping small patches of trees, but their use in large-area mapping has been limited. In this study, multi-temporal Satellite pour l'Observation de la Terre 5 (SPOT5)
... Resolution Geometrical data was pan-sharpened to 5 m resolution and used to map tree cover for the Australian state of New South Wales (NSW), an area of over 800,000 km 2 . Complete coverages of SPOT5 panchromatic and multispectral data over NSW were acquired during four consecutive summers (2008-2011) for a total of 1256 images. After pre-processing, the imagery was used to model foliage projective cover (FPC), a measure of tree canopy density commonly used in Australia. The multi-temporal imagery, FPC models and 26,579 training pixels were used in a binomial logistic regression model to estimate the probability of each pixel containing trees. The probability images were classified into a binary map of tree cover using local thresholds, and then visually edited to reduce errors. The final tree map was then attributed with the mean FPC value from the multi-temporal imagery. Validation of the binary map based on visually assessed high resolution reference imagery revealed an overall accuracy of 88% (˘0.51% standard error), while comparison against airborne lidar derived data also resulted in an overall accuracy of 88%. A preliminary assessment of the FPC map by comparing against 76 field measurements showed a very good agreement (r 2 = 0.90) with a root mean square error of 8.57%, although this may not be representative due to the opportunistic sampling design. The map represents a regionally consistent and locally relevant record of tree cover for NSW, and is already widely used for natural resource management in the state. Fragmentation degrades ecosystems, reducing biodiversity especially in the smallest and most isolated fragments      . Some declines in biodiversity are evident almost immediately after fragmentation, whereas others increase over time, with extinctions occurring decades or more after disturbances  . Managing and conserving trees in fragmented forests and other naturally heterogeneous landscapes such as open woodlands, requires maps that can accurately depict the landscape pattern. High spatial resolution imagery can allow small fragments such as thin corridors of trees, and scattered individual trees, to be mapped. Such detailed maps, however, are usually only produced for small regions  , and it remains a challenge to create consistent high resolution maps over large areas. Unlike freely available Landsat data, higher resolution satellite data is often prohibitively expensive for mapping large areas. Furthermore, processing large volumes of high resolution data is computationally intensive, especially when multi-temporal imagery is required to correct for data gaps due to cloud, shadow, and other artefacts [9, 10] . This article describes the production of a large-area tree cover map for a fragmented, heterogeneous landscape using high spatial resolution satellite imagery. The map covers the Australian state of New South Wales (NSW), and was required by the state government for natural resource monitoring at both small and large scales. The following sections outline recent advances in tree cover mapping, before describing the study area, and previous tree cover maps for NSW. Tree Cover Maps for Natural Resource Monitoring Land cover products are increasingly valuable inputs to a range of scientific studies and resource management activities  . Researchers have focussed on land cover classification [12-16] and monitoring change in vegetation cover [7, 9,     . These studies have been used to assess the extent of remaining natural forest, the location of threats to biodiversity as well as the effectiveness of existing protected-area networks. The challenge is to produce consistent maps across large areas while maintaining local relevance and utility  . The ability to produce global maps of forest loss using Landsat is relatively new [9,10,18]. The methods are similar to studies previously limited to continental scales [15, 19, 21, 22] . However, these large-area studies do not perform equally across all landscapes. For example, tree cover can be particularly difficult to quantify in grassy and semi-arid woodlands that feature naturally open canopies with large gaps between tree crowns. Tree cover in these areas is often underestimated by global maps, whereas mapping at regional scales allows classification schemes to be more finely tuned [16, 17,      . Producing tree cover maps over large areas of fragmented forest and open woodland remains a challenge. Many studies are focused on forested areas rather than examining all tree cover, resulting in maps that are sensitive to the definition of forest used. For example, estimates of Canada's net forest loss doubled when low and high tree cover strata were included, largely due to extensive burning in open boreal woodlands . Previous regional estimates of tree cover and change in Australia have also varied due to the definition of forest, with woodlands and open forest under-represented [28, 29] . The work presented here was designed to overcome the issues described above, and map the heterogeneous patterns of fragmented forest and open woodland through using high spatial resolution satellite data. The method benefits from combining automated data processing suitable for continental scale mapping with manual editing learnt from regional scale mapping. The result is a tree cover map that can be adapted to suit many definitions of forest, and can be used for natural resource management at small scales, such as state-wide reports, and large scales, such as local planning. We have defined tree cover as trees and shrubs taller than two metres that are visible at the resolution of the imagery used in the analysis (5 m). Study Area NSW is located on the central, eastern coast of Australia between 141˝E and 154˝E longitude and 28˝S and 38˝S latitude (Figure 1 ). It covers a large area (809,444 km 2 ), across a wide range of Remote Sens. 2016, 8, 515 3 of 23 climates (arid, temperate, subtropical, alpine) and has a variety of vegetation types such as shrublands, grasslands, woodlands and forests (Figure 1 ). The climatic gradients strongly influence the main vegetation patterns: shrublands and grasslands in the arid west grade through to woodlands and then to forest in the humid east; species composition grades from warmer, wetter, subtropical forest in the northeast to cooler, temperate forest in the southeast  . Remote Sens. 2016, 8, 515 3 of 22 climates (arid, temperate, subtropical, alpine) and has a variety of vegetation types such as shrublands, grasslands, woodlands and forests (Figure 1 ). The climatic gradients strongly influence the main vegetation patterns: shrublands and grasslands in the arid west grade through to woodlands and then to forest in the humid east; species composition grades from warmer, wetter, subtropical forest in the northeast to cooler, temperate forest in the southeast  .