Home > Research > Publications & Outputs > Random Forest classification of Mediterranean l...


Text available via DOI:

View graph of relations

Random Forest classification of Mediterranean land cover using multi-seasonal imagery and multi-seasonal texture

Research output: Contribution to journalJournal articlepeer-review

<mark>Journal publication date</mark>06/2012
<mark>Journal</mark>Remote Sensing of Environment
Number of pages15
Pages (from-to)93-107
Publication StatusPublished
Early online date17/02/12
<mark>Original language</mark>English


A Random Forest (RF) classifier was applied to spectral as well as mono- and multi-seasonal textural features extracted from Landsat TM imagery to increase the accuracy of land cover classification over a complex Mediterranean landscape, with a large number of land cover categories and low inter-class separability. Five different types of geostatistical textural measure for three different window sizes and three different lags were applied creating a total of 972 potential input variables. Madograms, rodograms and direct variograms for the univariate case and cross- and pseudo-cross variograms for the multivariate case, together with multi-seasonal spectral information, were used in a RF classifier to map the land cover types. The pseudo-cross and cross variograms were used specifically to incorporate important seasonal/temporal information. Incorporating multi-scale textural features into the RF models led to an increase in the overall index of 10.71% and, for the most accurate classification, the increase was up to 30% in some classes. The differences in the kappa coefficient for the textural classification models were evaluated statistically using a pairwise Z-test, revealing a significant increase in per-class classification accuracy compared to GLCM-based texture measures. The pseudo-cross variogram between the visible and near-infrared bands was the most important textural features for general classification, and the multi-seasonal pseudo-cross variogram had an outstanding importance for agricultural classes. Overall, the RF classifier applied to a reduced subset of input variables composed of the most informative textural features led to the highest accuracy. Highly reliable classification results were obtained when the 16 most important textural features calculated at single scales (window sizes) were selected and used in the classification. The proposed methodology significantly increased the classification accuracy achieved with a spectral maximum likelihood classifier (ML). The kappa values for the textural RF and ML classifications were equal to 0.92 and 0.83, respectively.