【作者】 张明罡；

【导师】 洪伟； 吴承祯；

【作者基本信息】 福建农林大学 ， 生态学， 2009， 硕士

【摘要】 土壤侵蚀是在水力、风力、冻融、重力等营力作用下,土壤、母质及其它地面组成物质被破坏、剥蚀、转运和沉积的全部过程。水土流失是全球性的环境和灾害问题,并且已经对人类的生存与发展构成了威胁。中国是世界上水土流失最严重的国家之一,我国东南红壤丘陵区,由于长期的不合理的开发利用,土壤侵蚀退化严重,水土流失面积达800 000km~2,其中,严重侵蚀地占165 000 km~2。东南红壤丘陵区是我国仅次于黄土高原的严重水土流失区,其中土壤侵蚀已成为引起土壤退化的重要原因。土壤侵蚀模型是研究土壤侵蚀学科的前沿领域和土壤侵蚀过程定量研究的有效手段。土壤侵蚀模型按建模的方法可分为经验模型和物理模型等,其中经验模型是根据已掌握的先验资料,采用统计分析的方法拟合、预测小流域土壤侵蚀量与降雨侵蚀力指标、植被覆盖率、土壤有机质含量、土壤粘粒含量、坡度、坡长之间的关系。土壤侵蚀过程中影响因子较多、过程错综复杂,以过程为基础的土壤侵蚀模型可操作性差,有些指标在大面积应用时难以获取。为了进一步提高经验模型拟合和预测的精度,克服经验模型固有的地域性强,外延精度差等不足,本文提出三种非线性土壤侵蚀建模方法,为土壤侵蚀建模提供了新的思路。利用福建省天宝岩保护区内毛竹林内所得的实测数据,以立地条件、林分密度、竹龄结构、经营措施一致的毛竹林不同坡度、不同垦复方式、不同施肥方式、不同清理方式、不同施肥处理、不同的笋收获方式为预报因子,建立了基于核函数主元分析的非线性岭回归毛竹林土壤侵蚀拟合和预测模型,取得了较好的模拟精度,因子用常规方法即可获得,因而可以用来预测不同耕作体系下毛竹林土壤侵蚀总量,为进一步研究毛竹林耕作体系的革新与新耕作方法的实施提供了参考依据。基于单变量高维指数非线性GWS (n,1)模型及其参数搜索、辨识原理,以复相关系数最大为判别准则、将多变量分为主变量和辅助变量,进而构建出多变量高维指数非线GWS ( n,m)模型, GWS (n,1)及GWS ( n,m)模型具有非周期、非线性、非正态振荡递减行为特征。各自变量的弹性系数呈振荡递减趋势并渐近与零,这是本文模型与常规模型的最重要区别。经过计算机编程,模型参数搜索、辨识较为简便易行。利用闽东南地区土壤侵蚀数据建立扩展模型,利用该模型对闽东南地区红壤小流域坡地土壤侵蚀进行预测相关系数达到0.986,精度明显高于USLE模型,并将该模型应用于水土保持耕作体系下毛竹林土壤侵蚀预报,精度达到0.968,为极显著水平。以总体离差平方和最小为判别准则,其中最优化算法是采用实数编码的自适应遗传算法,搜索出模型参数,模型的参数选取最大程度的避免了人为因素,不同与人工神经网络模型中繁琐而缓慢的“黑箱式寻优”,数据关系更加的透明化,在样本数量较多的情况下也具有很强的适应性,以复相关系数最大为判别准则、将多变量分为主变量和辅助变量,模型的精度也满足在实际工程中的应用,具有很强的指导意义。建立多元回归预测模型,并且要取得相当的逼近和预报精度,其前提是:影响因子与预报因子之间确切存在模型所假定方法的相关关系,可以是线性的或非线性的,这样才能根据样本的情况估计其中的参数,由于各个预测因子之间的相关关系并不是一致的线性或非线性,存在各种相关形式,因此采用一致线性或非线性形式建立的回归模型不能真实的反映回归关系,影响预报和逼近的精度,为了协调考虑这些不均衡的相关关系,引入加权的思想,使影响重要的关系在回归方程中所占的权重较大,从而对模型精度的改善起到一定的作用。投影寻踪结合约束优化遗传算法对雷竹园和茶园土壤侵蚀模数进行拟合和预测,取得了较为理想的效果,相关系数达到0.99用精度较高,原文献中采用线性回归和线性逐步回归的方法对土壤侵蚀进行预测,相关系数为0.90和0.84,精度明显较低,利用投影寻踪结合多目标约束优化遗传算法对土壤侵蚀模数进行拟合和预测能显著的提高模型的精度,并将该模型应用于毛竹林水土保持耕作体系下土壤侵蚀预测,取得了理想的模拟结果。同时,避免出现样本数量较多情况下产生的“维数祸根”、“精度与复杂度互不相容”等问题,模型建立之前无需对样本做任何假设,减少了模型的主观人为性,将高维数据映射到了低维,为在低维空间中研究高维数据提供了便利。更多还原

【Abstract】 Soil erosion is the process of soil and other composition materials be destroyed, denuded, taken away and deposition that caused by the power of water, wind, frozen and gravity. Soil erosion is a problem of environment and disaster all over the world, the problem has affected the existence and development of human kind. China is one of the countries that bearing the problem of soil erosion. The area of soil erosion has reached 800 000km~2 in red soil regions of southeast part of China, the most serious area has reached 165 000 km~2 .The red soil regions of southeast part of China has become the most serious soil erosion area except loess plateau of our country.The studies on modeling of soil erosion are the most effective method in the field of soil erosion. There two kinds of model of soil erosion: Empirical model and Physical model. Empirical model is using the former information by the method of statistics to forecast soil erosion and find out the relations of the Rainfall erosion index, vegetation cover, soil organic matter content, soil clay content, degree of slope, slope length. The process of soil erosion is complicated and has too many factors, the model of soil erosion based on process is hard to manipulate, some factors are difficult to get is the model is used in extensive areas, so it may have limited value. For the studies on modeling of soil erosion based on process is a new field, so the studies on empirical model play an important role in the study of soil erosion. So in such conditions empirical model is the most effective method. We aimed to enhance the empirical model fitting and prediction accuracy, to overcome lack of extension and poor accuracy, we use three methods of nonlinear of modeling, given three kinds of model for soil erosion forecast, provide new ideas for the modeling of soil erosion:According to the data collected from TianBaoYan Protected Areas FuJian Province, we use the different gradient, different way of cultivation, different way of fertilization, different way of clearing, different way of dealing with fertilization, different way of bamboo shoots harvest under the same conditions of local condition, forestry density, structure of the age, the way of cultivation as the factors of soil erosion forecast, established a model of nonlinear ridge regression based on kernel function principal component analysis, the model is effective and superior, factors can be got easily, so it can be used in soil erosion forecast under different way of cultivation, at the same time it also provide directions both to the renovation of traditional cultivation and new methods.Based on the univariate non-linear high-dimensional index model GWS (n,1), according to the largest multiple correlation coefficient, we divide variables into auxiliary variables and main variables, another model was established multivariate non-linear high-dimensional index model GWS ( n,m), the two models have a character of nonlinear, non-periodic, non-normal. We get a series of parameters easily; the model can be widely used. The data for modeling is collected from the southeast part of FuJian province, the correlation coefficient reached 0.986, accuracy was significantly higher than USLE model and the model was applied to soil erosion forecast under the system of soil and water conservation cultivation in Phyllostachy Heterocycla, the correlation coefficient reached 0.968, for a very significant level.we use optimization algorithm adaptive real-coded genetic algorithm to find out a series of parameters, we select the parameters avoid the man-made factors, it’s different from artificial network model in the“black box-style optimization”, the relationship between the data is more transparent, in the case of a large number of samples it also adaptive to a maximum correlation coefficient criterion , we divide variables into auxiliary variables and main variables, the model is accuracy to meet engineering applications in practical, it has a strong guiding significance.If we want to set up multiple regression prediction model, and to achieve considerable accuracy of approximation and prediction, we should confirmed that the impact factor and prediction factors is exist between the exact models is correct between assumes methods, which can be linear or non-linear. According to the estimate of samples to get a series of parameters, as the correlation between predictors is not a consistent relationship between the linear or nonlinear, the existence of a variety of related forms, so a consistent form of linear or non-linear regression model should not reflect the true regression relationship between the impact of forecasting and affect the accuracy of approximation, considering the uneven correlation between the introduction of thought-weighted, so that the impact of the important relations in the regression equation in a larger share of the weight, thus accuracy of the model can be improved. Projection Pursuit combined with multi-objective genetic algorithm used in soil erosion prediction have achieved better results and have higher accuracy in bamboo garden and tea garden the correlation coefficient reached 0.99, in the original literature, the use of linear regression and linear regression methods to predict soil erosion on a correlation coefficient of 0.90 and 0.84, significantly lower in accuracy, the model was applied to soil erosion forecast under the system of soil and water conservation cultivation in Phyllostachy Heterocycla stands to obtain the ideal simulation results. At the same time, it also avoid the situation "dimension curse" that have a large number of samples and the”precision and complexity can’t coexistence", the model was set up without any assumptions reduce the man-made factors , high dimensional data will be mapped to a low-dimensional, it provide facilitates to study high-dimensional data in low-dimensional space.更多还原