Incorporating Spatial Variability Measures in Land-cover Classification using Random Forest

Abstract

The spatial variability of remotely sensed image values provides important information about the arrangement of objects and their spatial relationships within the image. The characterisation of spatial variability in such images, for example, to measure of texture, is of great utility for the discrimination of land cover classes. To this end, the variogram, a function commonly applied in geostatistics, has been used widely to extract image texture for remotely sensed data classification.

The aim of this study was to assess the increase in accuracy that can be achieved by incorporating univariate and multivariate textural measures of Landsat TM imagery in classification models applied to large heterogeneous landscapes. Such landscapes which difficult to classify due to the large number of land cover categories and low inter-class separability. Madogram, rodogram and direct variogram for the univariate case, and cross- and pseudocross variograms for the multivariate one, together with multi-seasonal spectral information were used in a Random Forest classifier to map land cover types.

The addition of spatial variability into multi-seasonal Random Forest models leads to an increase in the overall accuracy of 8%, and to an increase in the Kappa index of 9%, respectively. The increase in per categories Kappa for the textural Random Forest model reached 30% for certain categories. This study demonstrates that the use of information on spatial variability produces a fundamental increase in per class classification accuracy of complex land-cover categories.