Background  Rocky desertification is one of the most important geo-ecological disasters in southwestern China. It causes land resources loss, ecosystem degradation, drought and water shortage, which seriously threatens the ecological balance, food security and the absence of large-scale return to poverty in southwest China. Accurate extraction of rock desertification information is crucial to the sustainable development of regional economy and society.

Methods  Aiming at the problems such as single temporal phase, poor timeliness and low accuracy of regional scale extraction results in the current rocky desertification information extraction, this study took Zhaotong city of Yunnan as an example by proposing an optimized classification method incorporating multi-features. Based on the preferential selection of samples and features, the multiple features such as spectra, indices, fractional vegetation cover, bedrock exposure rate, texture and topography were extracted using Sentinel-2 imagery and DEM data, and five classification schemes were constructed, as well as the extraction was completed using the random forest classification algorithm.

Results  1) When the Jeffries-Matusita (JM) distance algorithm was applied to evaluate separability of input features, the input features with the maximum average JM distance were BSI and Albedo, followed by TF₁ and slope, and the input features with the minimum average JM distance were B6 and B8. For rocky desertification land and other land cover types, slope, TF₁, BSI and Albedo had JM distance greater than 1.9, indicating a significant effect on the classification accuracy. 2) The importance of all input features was analyzed by the forest classification algorithm. The slope feature contributed the most to the classification accuracy, followed by the texture feature TF₁, NDVI and BSI, and the contribution of B4 and B6 bands in the spectral feature was relatively small. 3) In the case of the same number and distribution of sample points, compared with the other four classification schemes, the overall accuracy (OA) of the feature selection scheme obtained by using JM distance was 88.0%, and the Kappa coefficient was 0.85. The producer accuracy (PA) and user accuracy (UA) of rocky desertification land reached 91.2% and 83.8%, respectively. Finally, the rocky desertification land area of Zhaotong in 2020 was 2 820 km², accounting for 11.11% of the total land area of the region. The classification results were also in good agreement with the field survey area.

Conclusions  In this study, the input samples and characteristics are optimized by combining land use and land cover data and JM distance algorithm respectively, which effectively improves the phenomenon of misclassification, omission and large error in the fractured area of rocky desertification distribution in plateau mountainous areas. By the method proposed in this study, high classification accuracy at regional scale can be achieved, which provides reference for relevant departments to carry out rocky desertification prevention and monitoring.