多分类器组合支持下的山地森林覆盖类型分类与不确定性研究

The mountainous forests, as one of the most crucial ecological barriers, can not only keep the ecological equilibrium of the region, but also contain very abundant natural and economic resources, it can invigorate the rapid and healthy development of society and economic. A comprehensive understanding for the composition and structure of the mountainous forests can supply more scientific advices for the protection and utilization which can also ensure inclusive growth which is balanced and sustained. The forests are widely distributed and the topography of the mountainous region are always complicated which is hard for field investigation, but the highly developed remote sensing technology help us solve these problems in a more efficient way. However, the uncertainty problems of remote sensing classification are extensively existed, which imposed restrictions on the ground-object identification, thus finding out a new method which can combine the merits of different classifiers to decrease the uncertainty. The project will start with the discussion of the 5 different classifiers for the ability to identification ground objects, by the comparison of classification results, the classifiers with the most greatly difference will be integrated based on the Dampster-Shafer theory, finally the recognition mechanism for the mountainous forests types identification by 5 different single classifiers and integrated classifiers would be discussed and decided. The project supply algorithm support for mountainous forests, also provide theoretical basis for the forests resources monitoring and change detection.

森林作为重要的生态屏障，对于维护区域生态平衡具有重要的影响，对森林的组成和结构深入认知有助于对其进行科学保护和利用以使其平衡发展。实地勘测对于森林监测和保护具有重要的意义，但在由于复杂地形的影响，大部分山地森林的可及度较差，高分辨率遥感影像是监测山地森林资源的重要手段。但由于遥感成像中“不确定性”的广泛存在，对不同分类器识别地物的精度影响各不相同，因此寻找合适的方法将不同分类器的优点进行集成以减小分类器“不确定性”的影响成为遥感分类研究的主要目标之一。本项目拟从探讨不同分类器对地物识别的影响着手，选择差异性较大的分类器进行基于D-S证据理论的分类器集成研究，分析不同分类器及不同分类器组合方式对森林类型识别的影响机制，基于研究结果探讨证据理论是如何影响遥感“不确定性”。研究结果为复杂地形下的高山森林覆盖类型遥感识别研究提供算法支持，为山区森林资源监测提供理论基础。

山区地形条件和森林冠层结构的复杂导致森林植被覆盖类型遥感识别的精度较低，又考虑到不同分类算法对地物识别能力不一致等问题，本项目选择林芝市巴宜区为研究区，以Landsat OLI影像为遥感数据源，森林资源调查数据、样地实测数据和数字高程模型DEM为辅助数据，基于D-S证据理论对不同分类器进行组合以开展森林植被覆盖类型识别研究；在森林植被覆盖类型分类结果基础上，以数字高程模型DEM为支撑，对研究区复杂地形条件下的森林植被覆盖类型变化进行监测，以期从时空四维揭示复杂地形的森林植被覆盖类型变化，本研究为提升遥感图像分类技术在森林资源调查领域的深入拓展提供科学理论。结果表明对研究所选的三种分类结果差异性最大的分类器（最大似然分类器、马氏距离分类器和人工神经网络分类器）基于D-S证据理论对进行组合可以有效的提升遥感影像分类精度，同时也表明证据源的精度高低对组合分类器的分类结果有较大影响，可信度低的分类器会导致组合分类器的分类结果可信度较差，说明分类结果精度的不确定性会随着证据源的精度的提升而减小。此外，根据监督分类结果求解研究区范围内森林植被覆盖类型变化区和保持区的空间曲面面积，使得对地物的变化监测不再局限于二维平面，实现了对复杂山区的森林植被覆盖类型时空四维的变化监测，为森林资源监测提供科学支持。

内容获取失败，请点击重试