报告题目：Evaluating Digital Soil Maps by Their Patterns 

　　报 告 人：Professor David G. Rossiter（罗大伟教授） 

　　主 持 人：秦承志 研究员 

　　报告时间：2023年10月27日（周五）下午13:30-14:30 

　　地    点：中国科学院地理科学与资源研究所 D0301会议室 

　　报告人简介：My main research interest is in modern methods of soil resource inventory and the multi-purpose interpretation of soil geographic databases for both rural and urban applications. Both of these require quantitative methods including geostatistics, geographical analysis, statistical computing and modelling. I also work on numerical taxonomy of soils for objective classification, soil classification, soil survey methods, urban soil mapping and classification, and citizen science for soil mapping.  

　　Visiting Scientist, Chinese Academy of Sciences Nanjing Soil Science Institute； 

　　Guest Researcher, ISRIC-World Soil Information, Wageningen (NL)； 

　　Adjunct Professor, Section of Soil & Crop Sciences, Cornell University (USA)。 

　　《European Journal of Soil Science》Associate Editor，《Geoderma》、《Catena》等期刊编委。 

　　报告简介：Digital soil maps (DSM) are usually evaluated by point-wise "validation statistics". This evaluation has many problems: 

　　1.  It is based on a necessarily limited number of observations, far fewer than the number of predictions (grid cells, pixels). 

　　2.  The evaluation points are very rarely from an independent probability sample. 

　　3.  Cross-validation and data-splitting approaches rely on a biased point set. 

　　4.  Evidence has shown that widely different DSM approaches can result in maps with quite similar "validation statistics" but obviously different spatial patterns. 

　　5.  Soils are managed as units, not point-wise. 

　　6.  Land-surface models often rely on 2D or 3D connectivity between grid cells. 

　　7.  More than a century of fieldwork has shown that soils occur in more-or-less homogeneous patches of various sizes, not as isolated pedons. 

　　8.  DSM with obvious (visual) pattern differences often have quite similar “validation statistics”. 

　　Therefore I propose to evaluate DSM also by their spatial patterns, using well-developed techniques from landscape ecology and some new techniques on spatial segmentation.