Land cover mapping with patch-derived landscape indices

Automated-classification procedures of satellite imagery are mainly based on surface reflectance and generally ignore shape and size of landforms. On the other hand, quantitative landscape ecology has been focused on the patch concept as a landscape unit due to its relevance in the theory and practice of the conservation of species in human-modelled landscapes. The present paper explores how landscape metrics can introduce the component of spatial pattern of landscape elements to enhance land cover classification reliability. In particular, a method is proposed to extract patch-derived indices and to introduce them in a supervised classification of Landsat Thematic Mapper (TM) images as neo-channels. To extract patch-derived indices, an image segmentation method based on edge detection was used to define patches without an a priori knowledge of land cover classes. We calculated four patch indices: area, perimeter, shape index and fractal dimension. These indices were introduced in the classification to test the improvement of classification reliability. Well-known additional information (texture and topographic features) was also tested for comparison purposes. The method was tested for mapping land cover types in a mountainous region in the French Pyrenees, the Massif of Arize. The results of classification reliability allowed us to conclude that patch indices and topographic features significantly improved the discrimination of land cover classes. The combination of these additional information types by means of data fusion is useful for land cover classification purposes.