面向复杂地形的坡位K-means聚类划分研究

K-means Classifier for Automatic Slope Position Detection in Mountainous Areas

土壤植被研究建立在精准坡位划分的基础上。但现有的坡位大多采用手工划分的方式,存在着自动化程度低、划分精度不高且耗时较长等问题。本文提出一种顾及复杂地形的坡位自动划分算法,尝试采用机器学习K-means方法解决高海拔山区坡位划分的问题,并在山峰区域提取、聚类数确定、以及初始聚类中心选取等关键技术进行了算法的优化。为了验证算法的有效性,以云南省姚安县为研究区,运用提出的算法对研究区坡位进行自动划分,再采用Calinski-Harabasz聚类评价指标、调整兰德系数ARI和误差平方和SSE等一系列方法对坡位K-means聚类划分实验进行分析和评价。研究结果表明,利用该算法所生成的复杂地形坡位与研究区实测等高线相匹配。其次,再从姚安县规划风电场任选4个场址,比较13×13、25×25、37×37三种适宜窗口下坡位自动划分结果,结果表明选取25×25适宜窗口进行坡位划分可靠性最强。再者,计算的规划风电场内山脊、坡肩及背坡比例高达57.13%,也从一个侧面证实了利用该算法划分的坡位结果良好。

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.