The potential of support vector machine classification of land use and land cover using seasonality from MODIS satellite data

Abstract: With respect to climate change it is necessary to study land use and land cover (LULC) and their changes. LULC are related directly and indirectly to climatic changes such as rising temperatures that trigger earlier onset of vegetation growing seasons (IPCC 2007). Land surface phenology refers to the seasonal patterns of variation in vegetated land surfaces over large areas using satellite data (Reed et al. 2009). General variations observed from satellite may also be referred to as seasonality (Jönsson and Eklundh 2002, 2004). In this study, seasonality was modeled from normalized difference vegetation index timeseries derived from Moderate Resolution Imaging Spectro-Radiometer (MODIS) satellite data. Seasonality data contain valuable information about vegetation dynamics of LULC, such as the maximum of a season as well as the season start and end. The specific seasonality data signatures of LULC and may improve LULC classifications compared to multi-spectral satellite data approaches. Support vector machine classification (SVC) is a machine learning technique that does not require normal distributed input data. A normal distribution of seasonality data cannot be assumed. SVC is superior in comparison to traditional classification methods using multispectral satellite data (Tso and Mather 2009). Thus, it is feasible to test the potential of SVC separation of LULC using seasonality data. The most common linear and non-linear SVC methods recommended for satellite data were applied in this study. The chosen study area is located in southern Sweden, and its LULC classes are well documented by the latest CORINE land cover 2006 data. Thus, it is a good test area for validation of the performance of seasonality parameters for LULC classification using SVC. In this study, a SVC framework was developed and implemented that: (1) selects the most appropriate input seasonality data, (2) incorporates a direct acyclic graph for multiclassification and (3) validates the SVC outcomes with an accuracy assessment. The results of the four class SVC show moderate performances with overall accuracies between 61 - 64% and Kappa values ranging from 0.42 – 0.45. The accuracy differences between linear and non-linear SVC are marginal. However, there are potentials to improve the developed methodology, and thus the performance of SVC on seasonality data. In addition, the seasonality data should be tested with traditional parametric classifiers (i.e. maximum likelihood) in order to achieve valuable comparisons.