The slope position has generally been applied in a wide range of soil and vegetation studies. The slope position is manually classified in a long history into five types such as valley, footslope, backslope, slope shoulder, and ridge. It leads to issues of low automation, low precision and time consuming. This paper proposed a K-means algorithm of Machine Learning for clustering classification of slope position in mountainous areas. The performance improvements for the conventional K-means algorithm can be achieved by clustering number selection using the Calinski-Harabasz clustering evaluation index and by initial clustering centers finding using K-means++ in the context of slope position detection. The optimized K-means algorithm of a combination of peak area identification through the morphological white top hat transform function was applied into the automatic detection of slope position in Yao'an County, Yunnan Province based on 90 m×90 m SRTM DEM data. In order to validate this algorithm, a series of replicated experiments were carried out with different threshold values. Three accuracy measures of this algorithm such as Calinski-Harabasz clustering evaluation index, Adjusted Rand index and SSE can be estimated for these experiments. The results show that: (1) the best performance of this K-means algorithm is achieved with a clustering number k = 5; (2) this K-means algorithm is significantly better by using K-means++ to select the initial clustering centers than unoptimized selection; (3) the convergence of this K-means algorithm is the best if the iterations iter = 10,000. Furthermore, these results were obtained in a particular suitable window i.e. 25×25, and the window was compared to other two windows, that is, 13×13 and 37×37. An alternative statistical approach is the direct estimation of classification proportions of slope position for the study area, which can be achieved by evaluating point samples of backslope, slope shoulder, and ridge. Automatic mapping results in the planned wind farms are obtained up to 57.13%, which also indicates that the use of the proposed K-means algorithm may further enhance the potential of slope position detection. The advantages of our algorithm seem to lie in the help it gives for the development of automatic clustering classification of slope position as well as simple manipulation in spatial databases. Further improvements are needed in better performances by integrating fuzzy theory into this algorithm, suitable window selection by using the abruptshift analysis approach, as well as more topographic attributions such as slope, profile curvature and plan curvature, which will lead to the development of our algorithm.