基于可见/近红外光谱的中国北方苹果园土壤有机质含量检测方法

Methods for Detecting Soil Organic Matter Content of Apple Orchards in Northern China Based on Visible/Near-infrared Spectra

摘要:

土壤有机质 (Soil organic matter, SOM) 含量是衡量土壤肥力和生态质量的重要指标，然而国标规定的检测 SOM 含量的化学分析方法难以满足快速检测的需求。为了给 SOM 含量快速检测仪的研发提供基础，本研究采集了土壤类型复杂且空间差异显著的中国北方 10 个省区苹果园 760 份土壤样品。根据地理与气候特征，将样品划分为 4 个地区，即华北区、东北区、西北干旱区和西北边疆区。基于采集波长400~2450nm的土壤反射光谱以及根据国标方法测量的 SOM 含量，分析了 SOM 含量和地区差异对反射光谱的影响以及反射率与 SOM 含量的线性相关性，进而构建了检测 SOM 含量的偏最小二乘回归 (Partial least squares regression, PLSR)、支持向量回归 (Support vector regression, SVR) 和最小二乘-支持向量机 (Least squares-support vector machine, LS-SVM) 模型。结果表明，东北区土壤的 SOM 平均含量最高 (25.443 g/kg)，西北边疆区最低 (13.286 g/kg)；光谱反射率与 SOM 含量总体呈负相关；LS-SVM 模型对华北区和东北区 SOM 含量的检测性能最优，其剩余预测偏差 (Residual prediction deviation, RPD) 分别为 2.814 和 2.475，PLSR 模型对西北干旱区和西北边疆区 SOM 含量的检测效果最佳 (RPD 分别为 2.888 和 3.572)；LS-SVM 模型对检测 4 个地区混合样本 SOM 含量的性能最好 (RPD 为 2.864)。本文研究结果表明，构建适用于中国北方 10 省区苹果园 SOM 含量检测的通用模型是可行的，而构建各个地区的专用模型可以提高大部分地区 SOM 含量的检测精度。

Abstract:

The content of soil organic matter (SOM) is an important indicator for evaluating soil fertility and ecological quality. However, the chemical analysis methods prescribed by national standards for detecting SOM content are difficult to meet the demands of rapid detection. To provide a foundation for the development of a rapid SOM content detector, a total of 760 soil samples were collected from apple orchards in ten provinces in northern China with complex soil types and significant spatial differences. According to geographic and climatic characteristics, the samples were grouped into four regions: North China, Northeast, Northwest Arid, and Northwest Frontier. Soil reflectance spectra from 400~2450nm and SOM contents measured following the national standard method were used to examine the effects of SOM levels and regional differences on spectra and the linear correlations between reflectance and SOM. Partial least squares regression (PLSR), support vector regression (SVR), and least-squares support vector machine (LS-SVM) models were then constructed to predict SOM. The Northeast region had the highest mean SOM content (25.443 g/kg), while the Northwest Frontier had the lowest (13.286 g/kg). Soil reflectance showed an overall negative correlation with SOM. LS-SVM achieved the best prediction performance for North China and Northeast samples, with the residual prediction deviation (RPD) values of 2.814 and 2.475. PLSR performed best for the Northwest Arid and Northwest Frontier regions (RPD value of 2.888 and 3.572). For mixed samples from all four regions, LS-SVM provided the highest accuracy (RPD value of 2.864). These results indicated that building a universal SOM prediction model for apple orchard soils in the ten northern provinces of China was feasible, while building region-specific models was able to improve prediction accuracy for most regions.