Object‐Based Mapping of Karst Rocky Desertification using a Support Vector Machine

Accurate and cost-effective mapping of karst rocky desertification (KRD) is still a challenge at the regional and national scale. Visual interpretation has been utilised in the majority of studies, while an automated method based on pixel data has been nvestigated repeatedly. An object-based method coupling with support vector machine (SVM) was developed and tested using Enhanced Thematic Mapper Plus (ETM+) images from three selected counties (Liujiang, Changshun and Zhenyuan) with different karst landscapes in SW China. The method supports a strategy of defining a mapping unit. It combined ETM+ images and ancillary data including elevation, slope and Normalized Difference Vegetation Index images. A sequence of scale parameters estimation, image segmentation, training data sampling, SVM parameters tuning and object classification was performed to achieve the mapping. A quantitative and semi-automated approach was used to estimate scale parameters for segmenting an object at an optimal scale. We calculated the sum of area-weighted standard deviation (WS), rate of change for WS, local variance (LV) and rate of change for LV at each scale level, and the threshold of the aforementioned index that indicated the optimal segment level and merge level. The KRD classification results had overall accuracies of 85 50, 84 00 and 84 86 per cent for Liujiang, Changshun and Zhenyuan, respectively, and kappa coefficients are up to 0 8062, 0 7917 and 0 8083, respectively. This approach mapped six classes of KRD and offered a visually appealing presentation. Moreover, it proposed a conceptual and size-variable object from the classification standard of KRD. The results demonstrate that the application of our method provides an efficient approach for the mapping of KRD.