Comparison of function‐ and structure‐based schemes for classification of remotely sensed data

Abstract

The aim of this paper is to determine how classification‐scheme information content influences remote sensing classification accuracies. Two important informational constructs in environmental science are ‘process’ and ‘pattern’. In remote sensing these are analogous to ‘function’ and ‘structure’, ‘land use’ and ‘land cover’, or ‘informational’ and ‘spectral’ classes. The objective of this research was to test the hypothesis that structure‐based classes result in extraction of more accurate information than do function‐based classes. Two hierarchical, 19‐class schemes, one functional, the other structural, were developed for application with Satellite pour l'Observation de la Terre (SPOT) multispectral data for a watershed in North Sulawesi, Indonesia. Eight of the 19 classes were shared between the two schemes since these constituted equally valid functional and structural classes. Results indicate that there is no significant difference in classification accuracy between the functional and structural classifications as a whole (Khat = 82.2% and 84.9%, respectively). However, comparison of the two sub‐matrices associated with the 11 non‐shared classes showed significantly higher accuracies for the structural classes (Khat = 91.0%) than for the functional classes (Khat = 84.1%), thereby supporting the original hypothesis. Results demonstrate that careful consideration is required when developing function‐based classes for the extraction of thematic information from remote sensor data.

Acknowledgments

We thank the Collaborative Environmental Project in Indonesia (CEPI), under the Canadian International Development Agency (CIDA), for providing financial and logistical support for this study. Additional support was provided by the Natural Sciences and Engineering Research Council (NSERC) of Canada in the form of a research grant to Dr Treitz.