【作者】 徐丽华；

【导师】 谢德体； 魏朝富；

【作者基本信息】 西南大学 ， 土壤学， 2012， 博士

【摘要】 建立高效的土壤养分预测方法,提高土壤养分预测精度得到了越来越普遍的关注。随着信息技术的发展,尤其是地统计学、遥感、专家知识等新的方法和技术手段的应用,为土壤养分预测提供了新的数据与技术支持,大大提高了土壤养分预测的效率和精度。但是,土壤养分的空间变异性特征是尺度的函数,不同养分预测方法在不同尺度的研究中可能并不是普遍适用的；不同的研究区也往往具有不同的地理环境,特别是在西南丘陵山区,由于受到新构造运动的影响,地形割裂,多形成孤立的山丘和岗地,耕地破碎、分散,其地理环境相对复杂,新兴的土壤养分预测方法能否在地理环境相对复杂的较小空间范围内取得令人满意的精度是值得进一步探讨的。同时,对于研究方法的选择,在不同方法的比较的基础上确定最适合本地区的研究方法,往往更容易取得较可靠的预测结果。但目前的研究中,不同预测方法的比较多数集中于不同空间插值技术之间的比较,遥感和专家知识用于土壤养分预测多集中于大中尺度的研究,对于这些方法在小尺度研究上的适用性仍需进一步探索。本文在地理环境相对复杂的三峡库区王家沟小流域采集了111个土壤样本,在对土壤样本的理化性质和实验室反射光谱数据分析和测量的基础上,对空间插值、遥感技术、专家知识等土壤养分预测方法在较小尺度范围内的应用进行了探讨,并对不同方法进行了比较研究。主要研究结果包括：(1)基于空间插值的土壤养分预测。空间插值技术的不同以及模型内部参数的改变,都是影响土壤养分预测精度的重要因素。其中,逆距离加权和克里克方法对于土壤全氮、全磷、碱解氮、有效磷的预测都具有相近的均方根误差,但与逆距离加权插值方法相比,克里克插值在土壤全氮、全磷、碱解氮、有效磷的预测中都取得了较高的精度。在逆距离加权插值方法的应用中,土壤全氮、全磷、碱解氮、有效磷的最好插值结果并不对应相同指数和邻域数目,且邻域数目与指数等插值参数与数据的变异系数、偏度系数之间并不具有显著规律性。变异函数模型和邻域数目是决定克里克插值精度的重要因素。土壤全氮和全磷含量数据的最优变异函数模型是球形模型,碱解氮和有效磷含量数据的最优变异函数模型是指数模型。利用最优变异模型对土壤养分进行克里克插值发现,所有养分预测的最高精度出现的邻域并不相同,但过小的邻域往往不具有很高的插值精度。邻域数目与偏度系数和变异系数之间也未发现显著的规律。因此,在对养分进行空间插值之前,往往很难通过对土壤样本数据的统计分析来预先确定空间插值方法的内部参数。(2)基于高光谱遥感的土壤养分预测。由于受多种物质的综合影响,土壤的反射吸收特征可能并不明显。文中对土壤样本的反射光谱进行了去包络处理,将光谱曲线归一到一个一致的光谱背景,并在此基础上对光谱特征和特征波段进行选择,用偏最小二乘回归分析建立了紫色土和水稻土土壤养分的预测模型。其中,除了紫色土的土壤全磷和有效磷预测模型以外,其他模型得到的预测值与实测值值之间的相关系数都超过了0.5,而水稻土全氮含量的预测(?)度是最高的,其精度评价指标R达到了0.796。这说明,用高光谱方法预测养分含量具有一(?)的可行性。但通过不同类型土壤养分预测的交叉验证分析发现,紫色土和水稻土土壤预测模型的通用还具有一定的局限性,这可能是由于土壤类型的不同而造成。利用所获得的WorldView2和RapidEye卫星影像对将室内高光谱反演模型将多光谱卫星影像联合用于土壤养分制图的设想进行的验证发现,只有土壤全氮含量数据与RapidEye实验室模拟光谱在红波段的反射率表现出了较高的相关性,其相关系数为-0.53,土壤全磷含量数据与WorldView2和RapidEye的实验室模拟光谱的相关系数均低于0.4。但是,遥感影像上的实时光谱与土壤养分的相关系数的绝对值均低于0.3。因此,将实验室高光谱反演模型推广到多光谱影像,获得研究区的土壤养分图是十分困难的。(3)基于案例推理的土壤养分预测。从1m格网大小的DEM提取高程、坡度、沿等高线曲率、沿剖面曲率、地形湿度指数、坡位指数等环境要素数据,将这些数据和实测的土地利用类型数据作为环境因子,利用案例推理方法从土壤样本中获取土壤-环境关系知识,根据样本点土壤和案例点土壤之间环境组合的相似程度对土壤养分进行预测,进而表达土壤养分在空间上的详细变化。研究结果显示,林地、园地、水田和旱地土壤全氮、全磷、碱解氮含量的预测的平均相对误差都低于25%,而土壤有效磷含量的平均相对误差均高于80%。文中所用的基于案例推理的土壤养分预测方法的精度是与样本点的变异系数存在显著的相关关系,变异系数越小,土壤养分含量预测的精度越高。(4)土壤养分预测方法的比较分析。分析了采样方式、采样密度、采样数目发生变化时,基于空间插值、高光谱遥感、案例推理三种土壤养分预测方法精度变化,并对这些方法用于实现土壤养分从点到面,进行土壤养分制图的潜力和效果进行了评价。其中,上述三种方法得到的土壤有轰磷含量预测的总平均相对误差都超过了80%。且在多数情况下,案例推理方法都对土壤有效磷含量的预测取得了最好的精度,而空间插值方法的预测精度是最低。对于土壤全氮、全磷、碱解氮含量的预测,高光谱方法在不同的采样方式、采样密度和采样数目下都取得了最高的预测精度,空间插值方法的精度最低,案例推理方法的精度居中。由于数据获取的现状和方法自身的局限性,基于遥感的土壤养分预测方法实现土壤养分由点到面的转换相对较为困难,而空间插值方法和基于专家知识的土壤养分预测方法则比较容易。从空间详细度的比较可以看出,基于专家知识的土壤养分预测方法得到的土壤养分分布图具有更高的详细度,尤其是在没有采样点分布的区域,也具有较为细致的纹理,这个优点是空间插值方法比拟的。综上所述,文中通过基于空间插值、遥感、专家知识等方法在小尺度研究范围内的应用,为土壤养分预测方法在小尺度范围内的应用提供了新的案例支持；通过对复杂地理环境条件下土壤养分预测方法的比较,选择了具有较高预测精度、较好的制图潜力和制图效果的基于专家知识的方法作为本研究区最适宜的土壤养分预测方法,为更准确评价小尺度范围内土壤养分分布现状提供了合理的依据与技术支持。更多还原

【Abstract】 Attention is paid to set up high-prediction method of soil nutrients, as well as improve prediction precision.With the development of information technology, especially the utilization of geo-statistics, remote sensing and expert knowledge in soil nutrients prediction, some new methods produced, which improve the efficiency and precision. However, the soil nutrients spatial variability characterized by scaling function, one method may not be suitable to various scales. Additionally, geographical condition is not always same in different areas, especially for south-western hilly and mountainous areas. Affected by new tectonic movement, topology there changed to many small tiny pieces, which makes dispersed farmland and complex environment. Doubtfully, the new prediction method can obtain a high prediction precision in small spatial scope with such dispersed pixels. As to choose a researching method, it is easier to find a better one after comparison. Generally, most reported comparisons are among different spatial interpolations, little attention was paid to new prediction method and technology such as comparison among spatial interpolation, remote sensing and expert knowledge.Taking Wangjiagou small watershed of Three Gorges Reservoir Area as researching zone with complex geophical environment,111samples were collected. Based on the soil physicochemical properties, reflection spectrum analysis and measurement, various methods of soil nutrients prediction were applied in such small spatial scope:spatial interpolation, remote sensing and expert knowledge, comparisons were done after that. Details are as follows:(1) Soil nutrients prediction using spatial interpolation methods. The accuracy of interpolation methods for soil nutrients prediction is influenced greatly by interpolation methods or interpolation parameters. Discussed impacts of inverse distance weighted (IDW), Kriging and interpolation parameters on prediction accuracy. Results show that kriging and IDW methods gave similar RMSE values for soil nutrients prediction. Kriging produced better results than IDW for interpolating soil TN, TP, AN, and AP. In all uses of IDW, the most accurate estimates were yielded by difficult powers and difficult neighbourhoods. We found no significant correlation between interpolation parameters used for IDW and variation coefficients, skewness. Fitted variograms for soil TN and TP are spherical models. AN and AP in the topsoil are best represented by the exponential model.Use best variogram models to do kriging interpolation, the highest precision has various neighborhoods, but small neighborhoods cannot get a high precision. There is no correlation between neighborhoods and variation coefficients, skewness.Therefore, before doing the interpolation, it is difficult to determine the interpolation parameters by statistical analysis. (2) Prediction based on high-spectum remote sensing. Soil reflection and absorption is not significant because of different physical influence. Firstly, removed continuum from the reflectance spectra of soil samples, made them into a same spectrum background, then chose the characteristic wave bands, and produced the prediction models of purple soil and paddy soil, with the method of partial least squares regressive analysis. Results show that except the TP, AP model of purple soil, the others’correlation coefficients between their predicted values and measured values are over0.5, but precision of TN in paddy soil is highest, with R reaches0.796. The above indicates, it is reasonable to use hyper-spectrum method.After cross-validation of soil samples from two soil-types, the prediction model shows some constrains in the two soil, which may because the different soil types. Since hyper-spectrum data is expensive, it is hard to use the method to do soil nutrients mapping. Using WorldView2and RapidEye satellite images and high-spectum based prediction models to map soil nutrients, after that we found absolute value of correlation coefficient between TN and red band of RapidEye simulated is0.53, that between TP and simulated all bands of RapidEye, WorldView2is lower than0.4. Absolute value correlation coefficient between real-time spectra and soil nutrients is lower0.3. So that, the method has some difficulty in real utilization.(3) Soil nutrients prediction with case-based reasoning. Extracted elevation, slope, plan curvature, profile curvature, topographic position index and topographic wetness index from DEM with1m grid size, soil type data measured land use types data as environment factors, using case-based reasoning to predict soil nutrients based on similarity between samples and cases. Results show that, the average relative errors of TN, TP, AN are lower than25%, whereas, AP is higher than80%. In method of case-based reasoning, prediction precision and coefficients of variation are correlated dramatically:a lower coefficient of variation brings about a higher prediction precision.(4) Comparisons among different prediction method. Since precision is influenced by sampling method, density and number. Methods of spatial interpolation, remote sensing and expert knowledge were evaluated. Results show that, the whole average relative errors of AP are more than80%, in most cases, case-based reasoning for AP prediction obtain the fine precision, but Kriging gets the lowest one. As to TN, TP, AN, hyper-spectrum prediction has a higher precision, which follows by case-based reasoning, Kriging has the lowest. Since hyper-spectrum images and multispectral images data in the same period was not collected, we no obtained soil nutrients maps by hyper-spectrum prediction models, however, the other two methods can easily to do the mapping. Case-based reasoning can get spatial distributionof soil nutrients with more details, spatially in the areas without soil samples, which can not be reached by spatial interpolation.Above all, we validated soil nutrients prediction methods of spatial interpolation, remote sensing and expert knowledgein small spatial scope. By comparing the different prediction methods, case-based reasoning method that with highest precision and best mapping potential and effects is most suitable in researching area, it offers reasonable basis and technology support for assessing soil nutrients distribution in small spatial scope with complex geophical environment.更多还原