Arboriculture & Urban Forestry 42(6): November 2016 Pixel-based methods cannot provide high classifi- cation accuracy because of the spectral variability in urban areas. This limitation can be resolved by utilizing additional information about objects, such as spatial, textural, and color, in addition to spec- tral information. Hence, object-based classification is the optimal technique for urban tree detection. URBAN TREE SPECIES DETECTION Information on tree species is important for urban planning, disaster management, and ecological safe- ty. Accurate, reliable, and expressive measurements on the types of urban vegetation can help urban planners and researchers reach their targets (Iovan et al. 2008; Hao et al. 2011; Gong et al. 2013). The concept of classification of tree species was intro- duced in forestry when satellite and aerial imaging techniques were used to monitor forests (Gou- geon 1995). Numerous studies on the detection of tree species in forests are available (Immitzer et al. 2012), but research on urban areas remains limited. The limitations of methods involving different satellites or airborne sensors is one of the chal- lenges in studies on urban tree species. For instance, classical methods, such as MLC, can be applied on multispectral imaging, but these methods cannot be applied to hyperspectral data because of small training samples. Hence, other techniques, such as SAM (Wania and Weber 2007; Forzieri et al. 2013), linear spectral unmixing, and spectroscopic library matching are utilized to classify urban tree species by using hyperspectral data (Zhang and Qiu 2012). Forzieri et al. (2013) applied ML, SAM, and spec- tral information divergence on airborne hyperspec- tral data (i.e., multispectral infrared visible imaging spectrometer) to detect 10 urban tree species (i.e., herbaceous, heatland, arundo donax, poplar, oak, pine, Cupressus, spruce, willow, and olive), and ML presented the highest accuracy of up to 92.57%. This high accuracy could be attributed to the avail- ability of LiDAR data because other researchers used them as ancillary or main data to improve the accuracy of classifying urban tree species (Voss and Sugumaran 2008; Tooke et al. 2009; Hofle et al. 2012; Zhang and Qiu 2012; Tigges et al. 2013). Zhang and Qiu (2012) used LiDAR data to classify urban forest species based on tree crowns (i.e., crown-based species classification) because such data can address the limitation of tree 405 crown-shaded side, small tree crowns (might be seen as one object), and boundary of tree crowns, which may lead to mixed pixels. These authors developed a method based on hyperspectral data by combining the fundamental aspect of the neural network and fuzzy logic. They used the AGFLVQ algorithm to distinguish 20 urban tree species, and the result demonstrated higher clas- sification accuracy (approximately 68.8%) than that obtained using other hyperspectral methods, such as SAM (approximately 39.95%). The classi- fication accuracy is less than the accuracy shown in the study of Forzieri et al. in 2013, although the difference can be due to the number and types of tree species (Forzieri et al.: 10 species, Zhang and Qiu: 20 species). For instance, when evergreen and deciduous trees are considered, the use of the Gaussian fuzzy learning vector quantization (GFLVQ) method is unsuitable because the determination of at least two spectra should be used for deciduous trees. However, the basis of the GFLVQ algorithm is that all spe- cies have the same spectral signatures, and one spectral signature is sufficient for evergreen species. This finding reflects the limitation of GFLVQ, which can be solved using ancillary data, such as LiDAR data (Zhang and Qiu 2012). As multispectral data can be applied to dif- ferent classification methods, most studies on urban forest species were conducted via multi- spectral imaging because of the other limitations of hyperspectral data (e.g., high volume, cost, and time required). Appendix II shows a summary of the detection of urban forest species via remote sensing This and different review demonstrated that most classification methods. study areas involve non-tropical areas, which contain abundant evergreen and deciduous species. Dis- tinguishing evergreen from deciduous species by using their spectral signatures in spring or autumn is easier than differentiating tree species in a tropical area. As tree species present differ- ent spectral characteristics, spectral signatures are useful in each species. However, pixel-based classifications, such as ML, and MD with multi- spectral imaging but without any ancillary data, such as LiDAR, demonstrate low accuracy [e.g., Ismail and Jusoff (2004): ML approximately 61%]. By contrast, the complexity of the environment ©2016 International Society of Arboriculture
November 2016
| Title Name |
Pages |
Delete |
Url |
| Empty |
Search Text Block
Page #page_num
#doc_title
Hi $receivername|$receiveremail,
$sendername|$senderemail wrote these comments for you:
$message
$sendername|$senderemail would like for you to view the following digital edition.
Please click on the page below to be directed to the digital edition:
$thumbnail$pagenum
$link$pagenum
Your form submission was a success. You will be contacted by Washington Gas with follow-up information regarding your request.
This process might take longer please wait
