Mapping a specific class with an ensemble of classifiers

• Published in: International journal of remote sensing Vol. 28; no. 8; pp. 1733 - 1746
• Main Authors: Foody, G. M., Boyd, D. S., Sanchez-Hernandez, C.
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis 01.04.2007; Taylor and Francis
• Subjects: Animal, plant and microbial ecology; Applied geophysics; Biological and medical sciences; Earth sciences; Earth, ocean, space; Exact sciences and technology; Fundamental and applied biological sciences. Psychology; General aspects. Techniques; Internal geophysics; Teledetection and vegetation maps; England; United Kingdom; Great Britain; Western Europe; Quality class; discriminant analysis; maps; Europe; Space remote sensing; Landsat; neural networks; Multiple layer; accuracy; cartography; Classifier; Support vector machine; Decision analysis; classification; Information; Decision tree; conservation; Perceptron; Allocation; imagery; Procedure; refinement; spatial variations
• ISSN: 0143-1161; 1366-5901
• DOI: 10.1080/01431160600962566

Often, in remote sensing, interest is focused on just one of the many classes that are typically represented in the area covered by an image. Various binary classifiers may be used to separate this specific class of interest from all others. It can, however, be difficult to identify the most appropriate classifier in advance. The selection of classifier is also often further complicated by a desire for information on classification uncertainty to indicate the spatial variation in classification quality. Here, five classifiers (a discriminant analysis, decision tree, support vector machine, multi-layer perceptron, and radial basis function neural network) were used to map fenland, an important class for conservation activities, from Landsat ETM+data. The classifications derived ranged in accuracy from 81.2 to 96.8%. The outputs of the classifications were also combined using a simple voting procedure to determine class allocation. The accuracy of this ensemble approach was 95.6%. Although marginally, but insignificantly (at 95% level of confidence), less accurate than the most accurate individual classifier, it is difficult to specify the most appropriate classifier in advance. In addition, the ensemble approach yielded class-allocation uncertainty information that may be used to help post-classification refinement operations and later analyses. For example, as only a small proportion of cases were allocated with a high degree of uncertainty and these contained most of the mis-classifications, targeting such sites for fieldwork could be one simple and efficient means of increasing classification accuracy. Alternatively, the cases for which all five classifiers agreed on an allocation could be treated as being correctly labelled with a high degree of confidence.