Extraction Of Soil-environment Relationships From The Text For Predictive Soil Mapping

Soil,the material basis for human survival,is indispensable and non-renewable natural resource.The correct understanding of soil spatial distribution is the basis for the rational use of soil resource.Therefore,the spatial distribution of soil has important research and application value,and the acquisition of it has important research significance.Predictive soil mapping is the main method to obtain the spatial distribution of soil,and the acquisition of soil-environment relationship is a key issue for predictive soil mapping.In addition to soil samples,conventional soil maps,and experienced soil surveyors,text about soils(e.g.,soil survey reports)is also an important potential data source for extracting soil-environment relationship.If text data can be used as a kind of data source to extract soil-environment relationship,it can not only increase the chance to extract soil-environment relationship,but also be the only data source to acquire soil-environment relationship besides direct sampling,when in certain circumstance,such as,other data sources are not available.In order to extract quantitative soil-environment relation from textual materials,two issues need to be tackled:the first is the extraction and structuration of soil-environment relationship information from soil survey reports;the second is the acquisition of quantitative soil-environment relationship.To tackle the first issue,this thesis did as follows:firstly,designed a framework suitable for expressing soil-environment text information;secondly,expounded the construction methods of the domain knowledge base and corpus needed for information extraction;thirdly,analyzed and summarized the language description characteristics of soil information,environmental factor information and the characteristics of existing information extraction methods;and then,raised to use either a rule-based method or conditional random fields(CRFs)model based method for different types of information extraction,according to the variables.These types of variables include slope,elevation,annual mean temperature,etc.When the written text information is in uniform,the rule-based approach is used to extract information.But,when the information contained in written text is in diverse styles,the CRPs-based method is adopted.After that,finally,summarized the extraction results of different target variables,including landform and parent material,and filled these target variables into the corresponding soil-environment text information frame.To tackle the second issue,this thesis developed a method for mining quantitative soil-environment relationship based on soil-environment text information.Assuming that the distribution of soil types in geographic space corresponds to the combination of specific environmental factors in attribute space,the problem of acquisition soil-environment relationship is transformed into the acquisition of specific environmental factors combination and corresponding soil semantic information.The method can be divided into three steps.Firstly,the clustering method is used to obtain the combination clusters of environmental factors in each parent material division based on the soil-environment text information.Secondly,the similarity between the environmental factor combination cluster and the text information framework is calculated based on the cluster information and the quantitative environmental factor information in soil-environment text information framework,and then the semantic information of each cluster is determined.Finally,representative point sets are selected from the cluster of environmental factors,and the quantitative soil-environment relationship can be obtained by kernel density estimation method.The soil survey reports during the second soil census in China are selected as the experimental dataset,and the method proposed in this paper is applied and verified in Xuanzhou study area.Study results demonstrate that the rule-based method and CRFs-based method can effectively extract the target variable values when extracting structured soil-environment text information.For the uniformly written target variable values,such as elevation and slope,the precision is above 90%.And for target variable values contained in text written in diverse styles,such as landform and parent material,the precision is 88.62%and 84.15%respectively.The results also indicate that quantitative soil-environment relationship can be acquired based on soil-environment text information.The predictive soil mapping accuracy obtained by the proposed method can reach 69.77%,which is better than that of text filtering method(48.84%)and traditional soil map(51.16%).It can be concluded that the method proposed in this thesis can effectively obtain quantitative soil-environment relationship from text data,and illustrate that text data can be independently used as a data source to obtain soil-environment relationship.The proposed method further complement the existing methodologies of soil-environment relationships acquisition.