基于混合地理权重回归的土壤性质空间预测模型研究

Accurate spatial estimation of soil properties is great important for rational fertilization, environmental protection, and precision agriculture. It is, however, difficult to acquire detailed soil properties' data for a given region at low cost. Although a variety of geostatistical methods (e.g., kriging) are available, accurate prediction usually needs densely distributed sample data and requires the interpolated variable meets some constraints such as stationarity to ensure interpolation precision. Recently Geographically Weighted Regression (GWR) technique has been developed to explore the spatially non-stationary relationships and spatial autocorrelation in geographical fields. By adopting localized regression models, GWR has advantages in improving interpolation accuracy by incorporating information from various external environmental variables. As an extended method of GWR, the project aim to establish a mixed GWR model for accurate and reasonable estimation of soil properties. An integrated model will be adopted by combined local and global regression in a single one to estimate soil properties in space. To achieve the goal, we first put forward the mixed GWR model based on mathematical and statistical view, and then, precise expression and detailed computational process of constant & variable coefficients in mixed GWR model will be deduced based on parameter estimation approach. Soil properties will be gotten from soil samples, which is collected in the study area at a grid way. These properties will be used to test mixed GWR model. We will explore whether the model is feasible to interpolate soil properties, and also including model's precision, mapping effects and its applicability. The commonly used OLS, kriging, cokriging, regression kriging method are employed for the comparison purpose. We will also give a total report on the evaluation of soil properties distribution in the study area.

准确把握土壤理化性质的空间分布是合理施肥、保护环境的基础，也是实施精细农业的重要保障。地统计学作为众多学科进行空间预测的工具，也广泛用于土壤性质的预测，但该方法仍有明显缺陷。近些年在空间预测领域出现的地理权重回归方法，它充分利用变量与外部环境之间的相互关系，可提高预测精度，并能有效地解决地学领域空间非稳定性变量的预测问题。本项目将该方法作进一步拓展，采用局部与全局一体化模型用于预测连续土壤性质空间分布问题，以期提供一个对土壤属性精确预测的新方法。研究将首先确立混合地理权重模型的方案，通过参数估计推导出混合地理权重模型常系数和变系数的精确表达式，并给出计算过程。在研究区内，通过实地采样分析得到土壤的相关理化性质，然后用于混合地理权重模型模拟。再通过不同预测方法比较，探讨混合地理权重模型预测的精度水平、制图效果及适用范围。最后用该方法得到的结果，对研究区土壤肥力等性质的空间分布进行评价。

土壤性质空间的精确预测是实现精准农业、土壤质量评价及环境保护的基础。项目研究以构建精确预测土壤性质的混合地理权重模型为主要目标，通过模型构建、同类方法的预测比较的过程，研究了：1）同时能把空间稳定性参数及空间非稳定性因数在同一地理加权回归模型实现的途径，并研究了回归参数的解算方法；2）在选取的两个研究区（一个代表较少有人为干扰的区域，另一个代表人为干扰强烈的区域）内，进行样品的采集与资料收集，分析了区域土壤属性包括土壤有机质、土壤质地、有效P、全N、CEC、pH值、锌、铬、镉、铅、铜等多种痕量元素，利用上述模型分析它们的空间变异情况；3）分别用不同的空间插值方法（普通克里格法、回归克里格法、协同克里格法、普通线性回归法等）对上述的土壤性质进行内插预测，并对不同方法与地理加权回归方法进行综合对比，探讨它们的预测精度、制图效果、适用范围及各自的局限性；4）研究了定性因素（如土壤类型、土地利用、土壤侵蚀级别、坡度、坡向等）在参与模型过程中的定量途径与方法。通过上述研究，认为地理加权回归的方法能尽最大可能利用对土壤性质形成有重大影响的环境变量，对相关的土壤性质进行合理预测，对多数土壤性质能取得较好的回归精度，与制图效果，适合多数受环境因素影响较大的土壤属性的空间预测。