基于不确定性迭代优化的山地植被遥感制图

山地因其较高的异质性和特殊的环境特征给遥感科学及其应用带来了诸多问题和挑战。为实现山地植被信息的精准提取，本研究选择部分滇西北山地区域作为研究区开展方法实验，利用高分辨率遥感影像数据和数字高程模型，结合分区分层感知思想，提出一种基于不确定性理论的山地植被型组分类制图方法。首先结合地形对研究区影像进行多尺度分割制作图斑；然后根据图斑特征使用随机森林方法进行分类，将分类结果与对应类别样本间的相似性作为优化目标,并构建混合熵模型定量计算图斑推测类型的不确定性，据此进行针对性的样本补充和分类模型的迭代优化。实验总体分类精度达90.8%，较迭代前提升了29.4%,Kappa系数达到0.875。在高不确定性区域，该方法相比使用一次性补样和随机补样方法的分类结果，精度分别提高了17%和13%。研究结果表明，通过人机交互的方式，基于不确定性理论为样本库融入增量信息的迭代优化方法能够有效提高植被型组分类的精度，相较于传统的样本选择方法具有更高的效率和更低的不确定性。

Remote Sensing Mapping of Mountain Vegetation Via Uncertainty-based Iterative Optimization

Mountain area is an important part of terrestrial ecosystem and contains valuable ecological values. Due to its high heterogeneity and special environmental characteristics, there are many problems and challenges in remote sensing classification for mountainous areas. The traditional classification method based on vegetation index usually uses remote sensing data from a single source, which is effective in some scenarios, but severely limited in mountainous areas with fragmented landscape and complex topography. In order to achieve accurate mountain vegetation information, the mountainous areas in northwestern Yunnan were selected as research areas to carry out method experiments in this paper. This study used high resolution remote sensing image data and Digital Elevation Model(DEM), combined with the idea of zoning-stratified perception, and proposed a classification method for vegetation types in mountain areas based on uncertainty theory. Firstly, the images of the study area were segmented at multiple scales to make geo-patches under the constraints of the slope units, which were implemented by use of ridge lines and valley lines that were generated by hydrologic analysis based on DEM. Secondly, spectral, textural, and topographic features were selected for classification using random forest model. The experiment took the Mahalanobis distance as the similarity metric between the classification results and the samples of corresponding class as the optimization objective. Then the mixing entropy model was constructed to quantitatively calculate the uncertainty of speckle speculations caused by randomness and fuzziness, which depends on the membership degree of different vegetation types and the area proportion of different vegetation types. Finally, an automatic targeted sample supplement and iterative optimization of the model based on historical interpretation data, uncertainty theory, and similarity measurement were conducted. The model was updated accordingly every time the sample was supplemented. The iteration stopped when the Mahalanobis distance decreased to a convergence. This study also generated the variation trend of uncertainty in iteration and space. The overall classification accuracy of the experiment reached 90.8%, 29.4% higher than that before iteration, and the Kappa coefficient reached 0.875. In the high uncertainty region, the accuracy of this method was 17% and 13% higher than that of one-time and random sample supplement methods, respectively. The experimental results show that the method of iterative optimization, which integrates incremental information through human-computer interaction and imports high uncertainty and low confidence patches into the sample library, can effectively classify the vegetated mountain surface and has higher efficiency and lower uncertainty than the traditional sample selection methods.