Koch David E, Mohler Rhett L, Goodin Douglas G
Department of Geography, Kansas State University, Manhattan, KS 66506-2904, USA.
Geospat Health. 2007 Nov;2(1):15-28. doi: 10.4081/gh.2007.251.
Landscape epidemiology has made significant strides recently, driven in part by increasing availability of land cover data derived from remotely-sensed imagery. Using an example from a study of land cover effects on hantavirus dynamics at an Atlantic Forest site in eastern Paraguay, we demonstrate how automated classification methods can be used to stratify remotely-sensed land cover for studies of infectious disease dynamics. For this application, it was necessary to develop a scheme that could yield both land cover and land use data from the same classification. Hypothesizing that automated discrimination between classes would be more accurate using an object-based method compared to a per-pixel method, we used a single Landsat Enhanced Thematic Mapper+ (ETM+) image to classify land cover into eight classes using both per-pixel and object-based classification algorithms. Our results show that the object-based method achieves 84% overall accuracy, compared to only 43% using the per-pixel method. Producer's and user's accuracies for the object-based map were higher for every class compared to the per-pixel classification. The Kappa statistic was also significantly higher for the object-based classification. These results show the importance of using image information from domains beyond the spectral domain, and also illustrate the importance of object-based techniques for remote sensing applications in epidemiological studies.
景观流行病学最近取得了重大进展，部分原因是来自遥感影像的土地覆盖数据的可得性不断提高。以巴拉圭东部大西洋森林地区一项关于土地覆盖对汉坦病毒动态影响的研究为例，我们展示了如何使用自动分类方法对遥感土地覆盖进行分层，以研究传染病动态。对于此应用，有必要开发一种能从同一分类中生成土地覆盖和土地利用数据的方案。假设与逐像素方法相比，基于对象的方法在类间自动判别上会更准确，我们使用一幅Landsat增强型专题制图仪+(ETM+)影像，分别采用逐像素和基于对象的分类算法将土地覆盖分为八类。我们的结果表明，基于对象的方法总体精度达到84%，而逐像素方法仅为43%。与逐像素分类相比，基于对象的地图中各类别的生产者精度和用户精度都更高。基于对象的分类的卡帕统计量也显著更高。这些结果显示了利用光谱域以外领域的图像信息的重要性，也说明了基于对象的技术在流行病学研究遥感应用中的重要性。
