基于多源数据融合方法的中国1km土地覆盖分类制图

China Land Cover Classification at 1 km Spatial Resolution Based on a Multi-source Data Fusion Approach

精确的全球及区域土地覆盖数据是陆地表层过程研究的重要基础。在集成研究兴起和多种数据并存的背景下,利用多源信息融合技术进行大尺度土地覆盖制图具有重要的现实意义。证据理论清楚地表达了由于不确定和不完全信息所带来的对命题认识的"无知",能够确定相应的假设在目前的认知与知识状态下的确定、不确定和"无知"程度,是多源数据决策融合的重要方法。基于证据理论,将2000年中国1∶10万土地利用数据、中国植被图集(1∶100万)的植被型分类、中国1∶10万冰川图、中国1∶100万沼泽湿地图和MODIS2001年土地覆盖产品(MOD12Q1)进行了融合,最终基于最大信任度原则进行决策,产生了新的、IGBP分类系统的2000年中国土地覆盖数据。新的土地覆盖数据在保持了中国土地利用数据的总体精度的同时,补充了中国植被图中对植被类型及植被季相的信息,更新了中国湿地图,增加了中国冰川图最新信息,使分类系统更加通用。

Land cover plays a significant role inthe earth system science, which reflects the influence of human activities and environmental changes. In China, many land use/cover maps can be used in recent years. How to combine these date effectively to better produce land cover map is a key question. This paper presents a method to map land cover classification using Dempster-Shafer evidential reasoning （DS）. DS theory is a non-parametric classifier and can handle data from any number of sources at any scale of measurement and has an explicit mechanism for handling information uncertainty through the use of the concept of ignorance. The 1 km land use data of China in 2000 from China landuse database, the 1：1000000 vegetation map, the 1 km glacier map from 1：100000 glacier map, the 1 ： 1000000 wetland map and the MODIS land cover classification product （ MOD12Q1 ） as multisource of evidence to support each IGBP land cover class. These evidences are then combined using Dempster＇s Rule of combination. Finally, a new land cover map with higher accuracy is generated by a decision rule based on maximum belief function and could be representative of the Chinese land-cover status in 2000. A simple compare validation is taken and result shows that the accuracy of each land cover class of new land cover map has be strongly promoted relative to IGBP DISCover and MOD12Q1 land cover map especially for the accuracy of the croplands, urban, glacier, wetland and water bodies class. The evidence theory is can be used for fusing multi classification information. The fusion result is sensitive for the basic probability assignment of input data. The results will be validated further and the basic probability for multi-source input data will be assigned more scientific.