| Soil is a historical natural body being attached with the trait of highly inhomogeneous, spatial distribution of associated characteristic parameters has certain non-uniformity properties, as well as the scale effects accompanied spatially. Thus, a hot spot regarding researches of figuring out spatial variability and related variation processes in multi-scaled soil has sprung up in current irrigation hydrology. Generally, there exists variability in natural soil at both horizontal and vertical dimensions, plus, certain interactions are spatially detected in the soil properties of different soil layers. Given this, it is of great importance to explore the spatial variability and corresponding similitude relationships of soil properties in various soil layers, obtained results are capable of guiding the simulation and prediction of the variation patterns of soil properties in three-dimension space.Concerning the studies on spatial variation of soil hydraulic parameters, an essential approach is to establish conversion functions of hydraulic and physico-chemical parameters in obtaining wanted hydraulic parameters, however, corresponding dominant processes vary with the changed scales, making it worthy to study the subjects regarding how to apply a one scale-origin model to other scales through the scale conversion processes, to specify related accuracy and eliminate the uncertainties. Present study is supported by the National Natural Science Foundation of China (51069006) which utilized the Hetao Irrigation District in Inner Mongolia as the study area, and focused on the research and analysis of matters mentioned above.(1) The Bayesian neural network and BP neural network were applied to establish the model parameters of basic soil property parameters and characteristic curve of soil moisture, to work out the conversion functions between characteristic water contents, and to conduct the model validations and suitability comparisons. Both methodologies can provide satisfying training and forecasting results, herein, the training accuracy of BP model was higher than that of BNN model. However, the general accuracy of BNN model was superior to that of BP model, also, due to the fact that the forecasted values of BNN model varied in an interval, BNN model fitted better in revealing the spatial randomness and structural properties of soil property parameters.(2) The spatial variability of soil basic physical properties (clay, silt, sand and organic matter) and hydraulic parameters (saturated water content θs, van Genuchten model parameter a and n) in various soil layers (0-20cm,20-40cm,40-70cm,70-100cm) was analyzed at different scales (small scale:52.40km2, 1km×1km; medium scale: 1.24×103km2,4kmx4km; large scale:3.708x103km2,8kmx8km) using methods of Classical Statistics, Geostatistic and Multi Fractal techniques. Obtained results are as follow:Significant spatial autocorrelations were observed in the soil basic physical properties in different soil layers of different scales, the spatial distribution patterns were influenced by structural factors of parent material, climate, etc; remarkable multi-fractal traits were found throughout the research area, and the strongest spatial variability was seen in medium scale (apart from that of organic matter in soil layer depths of 0-20cm and 20-40cm). Besides, the Left Hook sign was illustrated in the multi-fractal spectrum curve (apart from that of organic matter in soil layer depths of 20-40cm and 40-70cm regarding the small scale), indicating that, in terms of spatial distribution, the dominating role was occupied by relatively larger data, associated probability distribution was comparatively larger as well.Generally, strong spatial autocorrelations were observed in hydraulic parameters in different soil layers of different scales, the spatial distribution patterns were impacted by structural factors of parent material, climate, soil types, etc; the Spatial distribution patterns in a parameter of van Genuchten model showed multi-fractal traits, however, non-coherent variability was seen in the multi-fractal spectrum width of all three scales. Insignificant multi-fractal traits were observed in the spatial distribution patterns of n parameter of van Genuchten model and saturated water content 9s, along with comparatively small multi-fractal spectrum width been found.(3) The Joint multi-fractal method was applied to study the correlation degree of spatial variability in basic physical characteristic parameters of surface layer (0-20cm) and other layers (20-40cm,40-70cm and 70-100cm), together with the conversion functions been built. In general, a decreasing order was followed in the relativity regarding spatial variability of all parameters in soil layers of 0-20cm,20-40cm,40-70cm and 70-100cm, also, corresponding relativity in small scale and large scale were higher than that of medium scale.The single variable function was adopted to establish the conversion functions of basic physical characteristic parameters in surface layer (0-20cm) and other layers (20-40cm,40-70cm and 70-100cm) of all three scales. The regression relations of all parameters in soil layers of 0-20cm and 20-40cm regarding all three scales were satisfying, the determination coefficient valued between 0.41 and 0.65, poor regression effects were detected in relation to soil layers of 40-70cm and 70-100cm, the determination coefficient varied from 0.038 to 0.401.(4) The medium scale based conversion functions of hydraulic parameters and basic physical characteristic parameters, hyper-spectrum and basic physical characteristic parameters were established through the application of methods of multiple regression, support vector and BP neural network, besides, related scale was shoved upward to the large one and shoved downward to the small one, with the applicability of associated scale transformations been assessed as well.The medium scale based hyper-spectrum and the inverse model of soil particle composition and organic matters performed well in other two scales. The relativity values of multiple regression method in other two scales ranged between 0.33 and 0.60, those of support vector varied from 0.41 to 0.52, those of BP neural network changed between 0.52 and 0.72, indicating the BP neural network-origin model had high applicability in other two scales. Plus, the general effects of particle composition were better than those of organic matters.Acceptable applicability performances in the conversion functions of medium scale based hydraulic parameters (saturated water content θs, van Genuchten model parameter a) and basic physical parameters (clay, silt, sand and organic matter) were observed in other two scales. The relativity values of multiple regression method in other two scales ranged between 0.535 and 0.944, those of support vector varied from 0.602 to 0.968, illustrating better applicability of support vector. Poor effects of modeling and model validating in the n parameter of van Genuchten model were seen. The scale conversion results of saturated water content θs appeared to be the best in all three parameters, followed by an Genuchten model parameter a and Genuchten model parameter n. |
PDF Full Text Request