| Landform is a key problem of surface environment, is the dominant element of natural region, obvious controls the regional distribution of surface water and energy, and indirect impact the succession and development on soil, vegetation, material migration and the ecological system. For the desertification area, topsoil thickness present different pattern on different landform. The research of landform and topsoil thickness provides important foundation for exploring soil, hydrological ecology, and surface environment, has important guiding significance for the practical activity, such as, water conservation, watershed management, evaluation and usage of groundwater resources. Ordos Plateau is located in the transition zone of the plateau and desert area, as an obvious ecological transitional zone and relatively independent natural unit, its ecological significance and ecological function are beyond all doubt. Exploring landform pattern and topsoil thickness distribution features can improve basic theory of ecological hydrology in the desertification area, and provide theoretical foundation and the scientific basis for construction of ecological hydrology, ecosystem health maintenance, and reconstruction of degraded ecosystem.In this study, we expounded landform classification methods, soil organic matter distribution and topsoil thickness distribution features in domestic and foreign research, then basis of field survey and laboratory experiment got a lot of surface features, including soil characteristic parameters, soil moisture, and soil organic matter and remote sensing data, such as elevation, slope, aspect, and NDVI. Based on that, build the landform classification model and topsoil thickness inversion model for northern Ordos Plateau of China by integrating remote sensing imageries and surface features, and analyze the influences of landform as a confounding variable on soil organic matter. The results were listed as follows:(1) In this paper, based on summarization of domestic and foreign research about landform classification method, and filed survey information, built landform classification theoretical parameter system that includes soil characteristic parameters and remote information. By comprehensive analysis, regression analysis and spatialization we got the data set that consist of D10, soil moisture, elevation, slope, aspect and NDVI, and the resolution of all of these data is30m, and this is the foundation for landform classification for the study area.(2) Using data collected at134landform sampling sites, each sample corresponding five independent variables:the sensitive soil curve’s descriptive parameter (D10), soil moisture, elevation, slope and NDVI. A set of Logistic regression equations were derived and model accuracy indicators system were build. The accuracy of the Logistic analysis was determined by using error matrices. The overall accuracy was51%in this case, indicated the number of sites where the class was correctly identified agree with51%of the total number of sites. The verification indicated these equations have moderate classification accuracy (Kappa coefficients K>43%). For a given grid (30m×30m), the Logistic equations were used to compute the landform type for this grid. In the study area M=2.6×107grid computation were was realized by a program written in Workbench(?)7.1.2IDL programming language. Finally, the landform map of the study area was generated from Logistic model.(3) For the200sampling sites, on the basis of appropriate improve of Random Forests program provided by Breiman and Cutler, nine different landform types considered as dependent variable, landform impact factors considered as independent variable, build the Random Forest landform classification model by integrating remote sensing imageries and surface features, and model accuracy indicators system. The out-of-bag (OOB) error is22.5%that obtained during the construction of Random Forests, it indicated the Random Forests model performed well. Furthermore, the Kappa coefficient is73%, and the averaged AUC (the area under the ROC curve) is0.99, these also shown the Random Forests model has good classification accuracy. For a given grid (30m×30m), the Random Forest model was used to compute the landform type for the grid. By a program written in Workbench(?)7.1.2IDL programming language, the landform map of the study area was generated from Logistic model.(4) Comparison of Random Forests and Logistic landform classification models, Kappa coefficient improved significantly from43%to73%, and overall accuracy increased from51%to77%. These demonstrated Random Forests landform classification model is better than Logistic model.(5) Landform classification is commonly done using topographic altitude and its extension information. In this case, we build Random Forests and Logistic landform classification models by integrating remote sensing imageries and surface features which are poorly documented in literatures, so it has some of the innovative. Especially Random Forests algorithm has good prediction ability even the training samples not much. And it won’t produce over fitting. User set parameters is less, and classification results are reliable.(6) According to the remote sensing sounding principle, using54field topsoil thickness data and seven band reflectance data of Landsat5TM, established a series of topsoil thickness inversion models, through the analysis of all the inversion results of the models, eventually got multiplied bands model (R=0.622) was the suitable model.(7) Through the comparison between observed and predicted values of multiplied bands model, it indicated that predicted values are closer to the observed values in10to140cm. However, beyond this range, the error was bigger, in0to10cm, predicted values were bigger than observed values and in110to300cm, predicted values were smaller.(8) Using bands combination of multispectral remote sensing to predicted topsoil thickness has a suitable range. In desertification area of northern Ordos Plateau, our conclusion is from10to140cm the predicted values preformed well. For other area, due to the different location, climate, sedimentary conditions, the physical properties of topsoil bottom, the depth of the detection range is different when using the multispectral remote sensing to predicted, and the suitable band combination or equation also will be different.(9) This study proposed topsoil thickness inversion is a new development for remote sensing sounding technology. Topsoil thickness inversion for desertification area has important scientific value for understand watershed underlaying surface conditions, build distributed hydrological model, and water resources system parameter simulation. For regional desertification control activities it also has important significance... |
PDF Full Text Request