Multisource Classification Using Support Vector Machines

Abstract

remote sensing classification are in vogue. In general, the mapping from these classifiers is based only on the spectral response of the classes. However, in areas particularly mountainous regions where there is large variation in the spectral response of classes due to high relief and shadow, mapping solely on the basis of spectral response may not be appropriate (Arora and Mathur, 2001) Moreover, information from an individual sensor may be incomplete, inconsistent, and imprecise (Rao and Arora, 2004; Simone et al., 2002) in some circumstances. Therefore, incorporation of additional or ancillary data sources in the process of remote sensing classification may result in better understanding and achievement of higher accuracy than utilizing spectral data from a remote sensing sensor alone. The ancillary data from various sources may be available in different forms and contexts, and at different frequencies, time, and spatial domains. Integration of data from different sources may also be referred to as image or data fusion (Pohl and van Genderen, 1998). Depending on the nature of data sources and methodology used, data fusion may be categorized as multi-source, multi-sensor, multi-temporal, multi-frequency, multi-polarization, or multi-resolution fusion (Arora and Mathur, 2001; Rao and Arora, 2004; Simone et al., 2002). The classification of remote sensing data along with data from other sources has generally been referred to as multisource classification. In the past, several studies (e.g., Benediktsson and Sveinsson, 2003; Bruzzone et al., 1999; Fitzgerald and Lees, 1994; Peddle et al., 1994) were conducted on multisource classification, and significant improvement in classification accuracy was achieved. These studies indicate that conventional parametric statistical classifiers have limitations in performing multisource classification. One major limitation of parametric classifiers is their inability to classify data at different measurement scales and units (Peddle et al., 1994). In these cases, the assumption that every dataset has the same statistical distribution may not be valid. Therefore, a non-parametric classifier may be appropriate, as it does not depend on the statistical distribution of the dataset. Examples of non-parametric classifiers for multisource classification in the remote sensing

Subpixel land-cover classification for improved urban area estimates using Landsat

“…IVM has been found to consistently outperform decision trees, artificial neural networks, and maximum likelihood algorithms (Watanachaturaporn, Arora, and Varshney 2008;Kotsiantis, Zaharakis, and Pintelas 2006;Huang, Davis, and Townshend 2002), with preferential (Braun, Weidner, and Hinz 2012) and comparable results to SVM (Roscher, Waske, and Forstner 2010). However, due to the heterogeneity of urban areas, it is important to calibrate these subpixel approaches against high spatial resolution data that capture the diverse characteristics found within urban environments (Lu, Moran, and Hetrick 2011).…”

Application of Support Vector Machines for Landuse Classification Using High-Resolution RapidEye Images: A Sensitivity Analysis

“…In case of data that has non-normal distribution (which is common with LULC data), the parametric classifiers may fail since the inability to resolve interclass confusion. This inability is the major limitation of parametric classifiers [Watanachaturaporn et al, 2008;Otukei and Blaschke, 2010;Pal, 2012]. Nonparametric classifiers like SVMs which do not rely on any assumptions for the class distributions of data, could overcome the aforementioned limitations of parametric classifiers [Kavzoglu and Colkesen, 2009;Mountrakis et al, 2011;Pal, 2012].…”

Enhanced land use/cover classification using support vector machines and fuzzy k-means clustering algorithms

“…Ancillary data have been used successfully to improve image classification, especially by including topographic measures (elevation and slope), normalized difference vegetation index (NDVI), or texture measures in the classification process additionally to the spectral information for separating features with similar spectral properties. 25,[35][36][37][38][39][40][41][42] NDVI has become a standard remote sensing product for ecological applications, 43 which has been widely applied for discriminating and interpreting mapped vegetation units. 44,45 NDVI was calculated from…”

“…Watanachaturaporn et al 25 have used the multisource classification with SVM. Different textural measures are a potential source of ancillary data and their benefits for LUC classification have been highlighted in studies using different techniques and classifiers.…”