Abstract: Karst areas have complex and irregular geomorphological features, which makes the accuracy of land use classification low. In this paper, Shanglin county in Nanning city is taken as the study area, and 33 feature variables are extracted and seven feature combination schemes are designed by combining multi-source data to explore the role of adding topography, texture, red-edge index and radar features on the extraction of land classes in karst areas. Combining the random forest OOB data error and recursive feature elimination method for feature optimisation, meanwhile introducing the third national land survey data to compare with the optimised classification results in order to evaluate its accuracy and reliability. The results of the study indicate that: ①Among the seven classification schemes, the traditional spectral features plus index features have the lowest classification accuracy, and the addition of topographic, texture, red-edge index and radar features can improve the classification accuracy, among which the texture features bring the most significant effect. ②The number of feature dimensions is reduced from 33 to 23 through feature optimisation, so that the classification accuracy reaches the highest, with an overall accuracy of 0.909 8, the overall accuracy is 0.909 8, and the Kappa coefficient is 0.884 9, which also reduces the complexity of the model and improves the computational efficiency. ③The classification results after feature selection are compared with the "three-tone data", and the overall accuracy is 0.852 5, which is in line with the actual situation of the study area. The classification method based on feature selection proposed in this paper can provide technical support and theoretical reference for the extraction of cover information in karst areas.

Key words: Karst landscape, land use, multi-source data, feature optimization, accuracy evaluation