Geo-scientific Knowledge Incorporated Into Remote Sensing Data For Automatic Soil Mapping Based On Soil Taxonomy In Hilly And Mountainous Region

Soil provides various raw products of agriculture, forestry and medicinal materials besides grain for the humanity. In addition, soil is the material source to maintain the biomass of ecosystem on the earth's surface. As the same as ocean, it is the necessary natural resource and environment for our living. With the development of the society, human's demand for soil resources information increases rapidly. Remote sensing technique has become a powerful tool for acquiring soil resources information due to its capability of covering vast area, rapidness and obtaining multi-temporal information repeatedly on the same area. At present, automatic soil classification based on remotely sensed data using computer has become the research focus for soil scientists, and is also the frontier in the field of soil remote sensing.Many studies in soil survey and classification by remote sensing have conducted by domestic and abroad scholars, but for hilly and mountainous region of subtropics which has moisture climate, plenteous rainfall, high vegetation cover and fragmentized ground surface, the problem of automatic soil classification using remotely sensed data has not been solved yet. Moreover, in our country, although the Chinese Soil Taxonomy, which is a classification system for soil types keeping up with the international normality, had been enacted in 1995, however, soil maps and databases covering large areas based on this system have not been produced up to date. Under such circumstances, in this dissertation we put forward a method to incorporate geo-scientific knowledge into remotely sensed data for soil recognition on the basis of the Chinese Soil Taxonomy in hilly and mountainous region of subtropics in our country.The process of our method consists of four steps. Firstly, image feature transformation is applied to remove vegetation and enhance soil spectral information. In the second step, soil classification using remotely sensed data is carried out by means of maximum likelihood classifier based on statistical decision-making model and radial basis function (RBF) neural network based on physiological vision model respectively. In the third step, the classification results of statistical decision-making classifiers and RBF network classifiers are integrated through multiple classifiers combination. Finally, as a post-classification step, geo-scientific experts' knowledge is added to correct the combination results of the third step by symbol logic reasoning. In the whole process, geo-scientific knowledge including DTM and some geo-scientists' knowledge is incorporated into remotely sensed data by two approaches: structured geo-scientific knowledge such as DTM data, regarded as a logical channel, is input into soil classification together with remotely sensed data; while unstructured geo-scientific experts' knowledge is integrated with remotely sensed data by D-S evidential reasoning.In this research, a case study using ETM+ data of Landsat-7 had been conducted to verify the proposed method above in Fengdu County of the Three Gorges Region. The major works and conclusions are summarized as following:(1) Through detailed analysis of soil conditions in the study area and the spectral information of all soil types on the ETM+ image, it is concluded that a majority of soil types in the study area could be recognized using remotely sensed image data by means of information enhancement, features transformation and feature selection, along with the help of some ancillary data. Simultaneously, it is also confirmed that ETM+ Band 6, namely thermal infrared band, is to conduce to improve soil classification accuracy, which is demonstrated by experimental results.(2) The linear spectral unmixing method is introduced to remove vegetation and enhance soil spectral information in subtropics of our country, where the percentage of vegetation cover is high and the spectra of vegetation and soil mix strongly. In this method, the fully constrained least squares (FCLS) algorithm for the retrieval of linear spectral mixture model with constraints is amended, and a new model is proposed to calculate pixel values after removing vegetation according to the proportion of vegetation endmember. The experimental results indicate that, ? the linear spectral unmixing method could effectively depress the spectral signal of vegetation, (2) being used for soil classification, the images facing-sun vegetation removed have the stronger discriminating capacity to stagnic anthrosols, although its overall accuracy is not higher than that of the raw images, and have better or similar classification accuracy compared to other several kind of commonly used feature transformations.(3) The problems related to remotely sensed image incorporating with structured geo-scientific knowledge, DTM data, for automatic soil classification are investigated. The results of maximum likelihood classifier show that it could improve classification accuracy remarkably by adding DTM data into soil remote sensing classification. The overall accuracies of image series of all feature transformations are increased by 10% above in average, and classification accuracies of most soil types are increased to different extent. Elevation in DTM data is discovered to play a most important role in soil recognition. The results of RBF network show that the performance of maximum likelihood classifier is superior to that of RBF network in classification accuracy, while the performance of the former is inferior to that of the latter in aspect of Kappa coefficient. When RBF network is used to discriminate natural objects like soils which have fragmentized distribution and for which training samples are difficult to select, if large samples input is restricted, it is proposed to divide large samples into many small sample areas, a set of pixel value averages of which are input into the network for training. This method not only lessens training samples, but also could maintain higher classification accuracy.(4) Based on elaboration of the D-S evidential theory, a fast integration of the abstract information of multiple classifiers' output had been implemented using two deduced rule of D-Sevidential combination formulation;which greatly reduced the computational complexity. The integrated results show that both overall accuracy and user' accuracy;as well as Kappa coefficient of the combination classifier are higher than those of the best single classifier. So the effect of soil classification map is improved in certain degree;because small patches classified incorrectly have been reduced obviously.(5) It is proposed to using the re-defined k and Psn in this dissertation as a guideline for the selection of classifiers to be combined. The experiment proved feasibility of the proposed method;and in general combining the classifiers with smaller k or Psn is able to obtain higher classification accuracy.(6) Soil classification using unstructured geo-scientific experts' knowledge integrated with remotely sensed data is studied;in which the output of multiple classifiers' combination is regarded as initial belief value belonging to each class for D-S evidential reasoning;combined with geo-scientific experts' knowledge and ancillary geospatial data. Under such condition that experts' knowledge about Soil Taxonomy is uncertain at the present time;it is proposed that the uncertainty of production rule expressing knowledge is denoted by both conditional probability (CF) and probability (CE) of production rule. And it is proposed how to transform rule into evidence;namely;the rules with the same conditions in the set of rules are regarded as one evidence supporting several propositions where the action part of each rule is the propositions supported by the evidence. The basic probability assignments of the propositions are defined as the product of CF and CE.The experimental results indicate that the overall accuracy of soil classification using unstructured geo-scientific experts' knowledge integrated with remotely sensed data is increased significantly;4.5% higher than that of multiple classifiers' combination as initial input;and reaches 71.2%;while the user accuracy of which increased by 7.5%;and the Kappa coefficient raised by near 10%. Small intermixing patches classified incorrectly are reduced to great extent;so the reliability of soil classification using remote sensing techniques is improved remarkably.