Abstract Soil is a three-dimensional volume with property variability in all three dimensions. In Digital Soil Mapping (DSM), the variation of soil properties down a profile is usually harmonised by the use of the equal-area spline depth function approach. Soil observations at various depth intervals are harmonised to pre-determined depth intervals. To create maps of soil at the defined depth intervals, 2.5D model produces maps of individual depth intervals separately. Those maps can be reconstructed to produce a continuous depth function for each predicted location. More recently, several studies propose that soil property at any depth can be mapped using a model incorporating depth along with spatial covariates as predictor variables, creating a ‘3D’ model. The aim of this study is to evaluate the proposition that soil properties can be predicted at any depth. This study compares the 2.5D model and 3D model in two areas. The first test is on a 1500 km2 area in Edgeroi, New South Wales (NSW), Australia, mapping soil organic carbon (SOC, %), carbon storage (kg m−2), pH (H2O), clay content (%), and cation exchange capacity (CEC, mg/kg) based on depth-interval observations. The second study area in the Lower Hunter Valley has SOC observations at every 2 cm increment from a 210 km2 area. 2.5D and 3D models were tested in both study areas using four machine learning techniques: Cubist regression tree, Quantile Regression Forest (QRF), Artificial Neural Network (ANN), and 3D Generalised Additive Model (GAM). Results show that, in terms of R2 and RMSE, 2.5D and 3D models using different machine learning models produce comparable results on the validation of depth interval observations. The 3D tree-based models produce “stepped” prediction of properties with depth. Results on the Hunter Valley area with point observations show that the 3D model cannot replicate field point observations. 3D soil mapping on point depth observation has lower accuracy and larger uncertainty compared to the 2.5D model. For future DSM studies, 3D soil mapping with depth as a covariate requires caution with respect to the prediction method and the requirements of the results.

中文翻译：

---

在 3D 中预测土壤特性：深度应该是协变量吗？

摘要 土壤是一个三维体，在三个维度上都具有属性可变性。在数字土壤测绘 (DSM) 中，沿剖面土壤特性的变化通常通过使用等面积样条深度函数方法来协调。不同深度间隔的土壤观测与预先确定的深度间隔相协调。为了以定义的深度间隔创建土壤图，2.5D 模型分别生成各个深度间隔的图。可以重建这些地图以生成每个预测位置的连续深度函数。最近，几项研究提出，可以使用将深度和空间协变量作为预测变量结合的模型来绘制任何深度的土壤特性，从而创建“3D”模型。本研究的目的是评估可以在任何深度预测土壤特性的命题。本研究在两个方面比较了 2.5D 模型和 3D 模型。第一次测试在澳大利亚新南威尔士州 (NSW) Edgeroi 的 1500 平方公里区域进行，绘制土壤有机碳 (SOC, %)、碳储存 (kg m-2)、pH (H2O)、粘土含量 (%)和基于深度间隔观察的阳离子交换容量（CEC，mg/kg）。下猎人谷的第二个研究区域从 210 平方公里的区域每增加 2 厘米就有一次 SOC 观测。使用四种机器学习技术在两个研究领域对 2.5D 和 3D 模型进行了测试：立体回归树、分位数回归森林 (QRF)、人工神经网络 (ANN) 和 3D 广义加性模型 (GAM)。结果表明，就 R2 和 RMSE 而言，2。使用不同机器学习模型的 5D 和 3D 模型在深度间隔观察的验证上产生可比较的结果。基于 3D 树的模型生成具有深度的属性的“阶梯式”预测。猎人谷地区的点观测结果表明 3D 模型无法复制现场点观测。与 2.5D 模型相比，点深度观测的 3D 土壤测绘精度较低，不确定性较大。对于未来的 DSM 研究，以深度为协变量的 3D 土壤制图需要谨慎对待预测方法和结果要求。猎人谷地区的点观测结果表明 3D 模型无法复制现场点观测。与 2.5D 模型相比，点深度观测的 3D 土壤测绘精度较低，不确定性较大。对于未来的 DSM 研究，以深度为协变量的 3D 土壤制图需要谨慎对待预测方法和结果要求。猎人谷地区的点观测结果表明 3D 模型无法复制现场点观测。与 2.5D 模型相比，点深度观测的 3D 土壤测绘精度较低，不确定性较大。对于未来的 DSM 研究，以深度为协变量的 3D 土壤制图需要谨慎对待预测方法和结果要求。