不同自然景观条件下土壤有机碳估算的主要环境因子研究

Obtaining the accurate amount and spatial distribution of soil organic carbon (SOC) is very important to soil fertility evaluation and global climate change modeling. Many methods for predictive mapping of SOC have been suggested all over the world. However, main environmental factors that influence the content of SOC under deferent natural landscape conditions were not effectively discriminated. So far existing studies mostly built models on global parameters or considered auxiliary variables globally, thus ignoring local variations of spatial correlations, which may inevitably result in large errors. Geographically weighted regression (GWR) has advantages over both conventional geostatistical methods and conventional multivariate linear regressions. Specifically, as a local regression method with spatially varied coefficients, GWR has remarkable advantages in local modeling and prediction error processing when using environmental factors as auxiliary variables. In this study, we will use GWR and GWRK (GWR kriging) to model the spatial distributions of SOC under different landscape situations so as to obtain the suitable sets of environmental impact factors and their influence strengths under different (e.g., geomorphological or land use) landscape conditions. In addition, the TF-IDF theory will also be used to determine the factor weights under deferent natural landscape conditions. Such an approach should represent a new idea for improving the prediction accuracy of SOC. This project will explore the follow aspects: (1) how to determine the main factors and their weights under deferent natural landscape conditions; (2) how to build the best fitting model of SOC to deferent sets of environmental factors; (3) the applicable scope and augmentability of the prediction datasets of SOC; and other related issues. These study results will be beneficial to improving the prediction accuracy of SOC, and will also provide an effective solution to more accurately estimating soil organic carbon storages at regional scales.

准确获取土壤有机碳含量及空间分布特征，对于探清土壤肥力状况和全球气候变化建模具有重要意义。国内外提出了许多土壤有机碳预测模拟方法，但由于没有区分不同自然景观条件下影响土壤有机碳含量的环境因子，建模时也只考虑了全局变量而忽略了局部环境因子，导致结果误差较大。地理加权回归模型融合了地统计学和多元回归模型的优点，尤其在结合环境因子作为辅助变量进行局部模型构建和预测残差处理方面优势明显；同时借助TF-IDF理论确定不同景观条件下土壤有机碳影响因子权重值，应是提高土壤有机碳含量预测精度的新思路。本项目将探索：（1）针对不同自然景观条件，如何确定土壤有机碳主导环境因子及其权重；（2）针对不同环境因子，如何构建最佳土壤有机碳预测拟合模型；（3）不同自然景观条件下，土壤有机碳预测数据集的适用条件及扩展性；等有关问题。这些研究有利于提高土壤有机碳含量预测精度，为更准确估算区域土壤有机碳储量提供有效解决途径。

土壤有机碳是土壤中较为活跃的组分，同时也是地球表层系统中最大的碳库之一，估算其含量对评价土壤健康、农作物产量和生态系统服务功能具有重要意义。土壤有机碳成因复杂且在不同空间范围内具有不同程度的空间变异性，加上多种自然环境和人为因素的影响，难以对其实现准确快速的预测。为此，本项目首先采用Pearson相关系数、地理探测器及改进的TF-IDF算法从土壤有机碳环境因子预选指标集中选择主要环境因子，并通过对比选取结果及原理差异确定最佳因子筛选方法。在此基础上，基于土壤-景观定量模型对土壤有机碳及其主要环境因子间分别建立空间回归拟合模型（多元线性回归、空间自回归和地理加权回归等）以及空间区域插值模型（普通克里格、回归克里格和地理加权回归克里格等），综合评价精度指标优劣，判定最适宜进行土壤有机碳预测的拟合模型。最终尝试将预测模型推广到不同自然景观条件或不同尺度区域中，确定其适用范围与条件。经过相关研究，获得以下成果：（1）不同环境因子筛选方法选取出的土壤有机碳主要环境因子大致相同，只是按贡献值大小的排序略有不同，但改进的TF-IDF算法能真正考虑到土壤样本的空间位置及不同景观间的环境因子差异，是获取土壤有机碳主要环境因子更为可靠的手段；（2）不同自然景观条件下的土壤有机碳空间分布存在显著环境梯度差异，且影响其含量变化的主要环境因子随景观类型变化发生改变，为缩小实测值与预测值间拟合误差，有必要对研究区中的杂糅景观进行明确划分；（3）在不同尺度或自然景观条件下进行土壤有机碳预测的研究结果均表明，地理加权回归克里格目前是土壤-景观定量模型中的最佳拟合模型，它既可依据距离衰退公式在不同地理位置上赋予环境因子以变化系数，又能揭示可能被空间非平稳性所掩盖的一些局部变化，反映出更加真实的土壤属性变异情况，为准确模拟土壤有机碳含量空间分布情况和提高数字土壤制图精度提供参考。