基于图像处理和SVR的土壤容重与土壤孔隙度预测

土壤容重和土壤孔隙度是衡量土壤结构的重要参数。传统的土壤容重、土壤孔隙度获取方法费时费力，且大多数预测模型的输入变量获取难度较高。该研究利用土壤粗糙度、土壤阻力与土壤容重的相关关系，以土壤表面图像的颜色参数和纹理参数表征土壤粗糙度，同使用车载式土壤阻力测量系统获得的土壤阻力一起，从信息融合的角度构建了支持向量机回归（Support Vector Regression，SVR）土壤容重预测模型和SVR土壤孔隙度预测模型。图像处理使用HSV颜色空间进行阈值分割，得到HSV颜色参数，纹理参数使用灰度共生矩阵的能量、熵、对比度和逆方差。使用主成分分析对颜色参数和纹理参数进行主成分提取。将SVR模型的预测结果同环刀法测得的标准值进行相关性分析，决定系数R2达到了0.867。土壤孔隙度SVR预测模型决定系数R2达到了0.743。在相同的运行环境下，将SVR模型与决策树回归模型结果做了对比，决策树回归对土壤容重和土壤孔隙度的预测精度R2分别为0.734和0.690，验证得到SVR预测模型具有更好的预测精度。研究可为节省试验成本，以及快速、有效预测土壤容重和土壤孔隙度提供方法和参考。

Predicting bulk density and porosity of soil using image processing and support vector regression

Abstract: Soil structure is essential for plant development and moisture balance, generally representing the spatial heterogeneity of different components or properties of soil. In this case, bulk density and porosity of soil are important parameters to evaluate the soil structure. In the traditional measurement, the ring knife is normally used to measure soil bulk density. But this commonly-used measurement requires multiple instruments, such as ring cutters, aluminum boxes, and drying boxes, although the measured data is accurate to serve as a standard requirement. Particularly, the whole process is time-consuming and labor-intensive, unsuitable for the rapid and accurate measurement of soil bulk density in a large range of farmland in recent years. Therefore, it is highly urgent to explore a convenient, efficient, and indirect measurement of soil bulk density, especially for the input variables for most prediction models in precision agriculture. In this study, prediction models of soil bulk density and porosity were constructed with the soil resistance using image processing and Support Vector Regression (SVR). The color and texture parameters of the soil surface image were also used to characterize the soil roughness, according to the correlation between roughness, resistance, and bulk density of soil. A measuring system was developed to mount a vehicle for soil resistance. In image processing, HSV color space was used for the threshold segmentation, while the first-order distance, second-order moment, and third-order moment of HSV three components were taken as color parameters. The specific texture parameters included the energy, entropy, contrast, and inverse variance of the gray-level co-occurrence matrix. Principal component analysis was used to extract the principal components of color and texture parameters for the non-correlation between the input parameters. The correlation analysis was then made between the prediction of the SVR model and the standard value measured by the ring knife. Specifically, the coefficient of determination R2 of the SVR model reached 0.867 for the prediction of soil bulk density, the coefficient of determination R2 of decision tree regression model reached 0.734 for the prediction of soil bulk density, and the SVR model root mean square error was 0.001 g/cm3, indicating better performance than that of decision tree regression. Nevertheless, the calculation time took 6.810 s, about 4.7 s longer than the 2.153 s calculation time of decision tree regression. In soil porosity, the coefficient of determination R2 of the SVR model was 0.743, and the root mean square error was 2.284. The coefficient of determination R2 was 0.690 for the decision tree regression model, the root mean square error was 3.345. The calculation time of the SVR model was 3.144 s, less than the duration of the decision tree regression model at 4.302 s. It demonstrated that the SVR model can widely be expected to predict soil bulk density and porosity using color and texture parameters combined with soil resistance as input variables. In the case of small and medium-sized data samples, SVR model can achieve good prediction results. The finding can provide a sound reference for the rapid and effective prediction of bulk density and porosity in soil.