Abstract:
Land cover classification using remotely sensed data is a critical first step in large-scale environmental monitoring, resource management and regional planning. The classification task is made difficult by severe atmospheric scattering and absorption, seasonal variation, spatial dependence, complex surface dynamics and geometries, and large intra-class variability. Most of the recent research effort in land cover classification has gone into the development of increasingly robust and accurate (and also increasingly complex) classifiers by constructing–often in an ad hoc manner–multispectral, multitemporal, multisource classifiers using modern machine learning techniques such as artificial neural networks, fuzzy-sets, and expert systems. However, the focus has always been (almost exclusively) on increasing the classification accuracy of newly developed classifiers. We would of course like to perform land cover classification (i) as accurately as possible, but also (ii) as quickly as possible. Unfortunately there exists a tradeoff between these two requirements, since the faster we must make a decision, the lower we expect our classification accuracy to be, and conversely, a higher classification accuracy typically requires that we observe more samples (i.e., we must wait longer for a decision). Sequential analysis provides an attractive (indeed an optimal) solution to handling this tradeoff between the classification accuracy and the detection delay–and it is the aim of this study to apply sequential analysis to the land cover classification task. Furthermore, this study deals exclusively with the binary classification of coarse resolution MODIS time series data in the Gauteng region in South Africa, and more specifically, the task of discriminating between residential areas and vegetation is considered.
