Digital soil mapping of several soil properties: Forest Hill Agricultural Research Station

A soil survey was conducted at CSIRO’s “Forest Hill” research farm to support scientific research, land management, infrastructure planning, and development activities. \n\nThis dataset complements Zund et al. (2024) that describes the main soil types and describes their morphological, chemical, and physical properties across the research farm.\n\nThis dataset contains digital soil mapping outputs for a range of agronomically relevant attributes, including soil texture, whole-soil bulk density, pH, electrical conductivity, exchangeable cations, and total soil carbon and nitrogen. In this context, comprehensive digital soil mapping refers to three-dimensional (3D) mapping using model-based approaches, with explicit quantification of prediction uncertainty. \n\nSpatial predictions (and associated uncertainties) were made on a 10 m × 10 m raster grid. \n\nZund, Peter; Cocks, Brett; & Malone, Brendan (2024): Forest Hill Agricultural Research Station Soil Map. v1. CSIRO. Data Collection. https://doi.org/10.25919/xsdz-nj28\nLineage: Digital Soil Mapping Workflow\nThe development of the digital soil mapping outputs followed a structured and systematic workflow, comprising the following key steps:\n\n1. Data-Informed Site Selection\nSoil core sampling locations were strategically selected to capture the maximum possible spatial variability in soil properties across the landscape.\n\n2. Soil Survey and Coring\nField surveys were conducted to extract intact soil cores for detailed analysis.\n\n3. Proximal Soil Sensing and Laboratory Analysis\nEach soil core was scanned using visible–near infrared (vis-NIR) spectroscopy and gamma-ray attenuation. Targeted subsampling was then undertaken for laboratory wet chemistry analysis. The resulting analytical data, combined with the spectral responses, were used to calibrate soil inference models capable of generating full-profile characterisations across all cores.\n\n4. Compilation of Environmental Covariates\nA suite of gridded environmental covariate layers was assembled to support the spatial modelling process. These covariates represent a diverse range of soil-forming factors.\n\n5. Digital Soil Mapping\nCalibrated soil profile data were integrated with the environmental covariates to develop spatial prediction models tailored to each soil attribute.\n\nThis workflow aligns with the approach described in Malone et al. (2022). Please see accompanying report for detailed steps and treatment of the data and modelling processes. \n\n\nOrganisation of Outputs\nFor each soil attribute, the outputs are organised into a folder containing three sub-folders:\n\nmapsout:\nContains visualisations of the 50th percentile (median) predictions for each specified depth.\n\nmodel_diogs:\nIncludes model diagnostics and performance metrics (from both calibration and testing sets) for each bootstrap iteration. Reported metrics include:\n\n- Coefficient of determination (R²)\n\n- Lin’s concordance correlation coefficient (CCC)\n\n- Mean squared prediction error (MSE)\n\n- Root mean squared prediction error (RMSE)\n\n- Mean prediction error (bias)\n\nrasters:\nContains GeoTIFF raster files of the mapped predictions and associated uncertainty. These are structured by depth and output type.\n\nDepth Convention and Output Types\nPredictions are generated for standard depth intervals from the soil surface to 180 cm. The file naming convention denotes these as d1 to d10, corresponding to the following depth slices:\n\nd1: 0–10 cm\n\nd2: 10–20 cm\n\nd3: 20–40 cm\n\nd4: 40–60 cm\n\nd5: 60–80 cm\n\nd6: 80–100 cm\n\nd7: 100–120 cm\n\nd8: 120–140 cm\n\nd9: 140–160 cm\n\nd10: 160–180 cm\n\n\nEach depth slice includes three prediction types:\n\nlower_percentile: 5th percentile of bootstrap predictions (lower confidence bound)\n\n50th_percentile: Median (central estimate)\n\nupper_percentile: 95th percentile (upper confidence bound)\n\nThese represent uncertainty bounds around the modelled predictions at each grid cell.\n\nReference\nMalone B, Stockmann U, Glover M, McLachlan G, Engelhardt S, Tuomi S (2022). Digital soil survey and mapping underpinning inherent and dynamic soil attribute condition assessments. Soil Security, 6, 100048.本数据集依托CSIRO旗下"Forest Hill"试验农场开展土壤调查，旨在为科学研究、土地管理、基础设施规划及开发活动提供支撑。 本数据集是对Zund等人（2024）研究的补充，该研究阐述了该试验农场内的主要土壤类型及其形态、化学与物理特性。 本数据集包含一系列农艺相关属性的数字土壤制图（digital soil mapping）成果，涵盖土壤质地、原状土壤容重、pH值、电导率、交换性阳离子以及土壤总碳与总氮。本研究中的广义数字土壤制图指采用基于模型的方法开展三维（3D）制图，并对预测不确定性进行明确量化。 空间预测（及相关不确定性）基于10 m × 10 m的栅格网格生成。 Zund, Peter; Cocks, Brett; & Malone, Brendan (2024): Forest Hill Agricultural Research Station Soil Map. v1. CSIRO. 数据采集. https://doi.org/10.25919/xsdz-nj28 数据谱系：数字土壤制图工作流 数字土壤制图成果的开发遵循结构化且系统化的工作流，包含以下关键步骤： 1. 基于数据的样点选择 土壤岩心采样点位经过策略性选取，以最大化覆盖研究区土壤属性的空间异质性。 2. 土壤调查与岩心采集 开展野外调查以获取完整土壤岩心用于后续详细分析。 3. 近地土壤传感与实验室分析 对每一根土壤岩心分别开展可见-近红外（visible–near infrared, vis-NIR）光谱扫描与伽马射线衰减（gamma-ray attenuation）检测。随后针对目标子样本进行实验室湿化学分析。将所得分析数据与光谱响应相结合，用于校准土壤推断模型，以实现所有岩心的全剖面特征刻画。 4. 环境协变量汇编 组装一系列栅格化环境协变量图层，以支撑空间建模流程。这些协变量涵盖了多样化的土壤形成因子。 5. 数字土壤制图 将校准后的土壤剖面数据与环境协变量相结合，为每一种土壤属性构建专属的空间预测模型。 本工作流与Malone等人（2022）所述方法保持一致。详细的数据处理与建模流程步骤请参阅随附报告。 成果组织形式 针对每一种土壤属性，其成果均收纳于一个包含三个子文件夹的目录中： mapsout： 包含各指定深度层50%分位数（中位数）预测结果的可视化成果。 model_diogs： 包含各自助法（bootstrap）迭代的模型诊断结果与性能指标（涵盖校准集与测试集）。所报告的指标包括： - 决定系数（Coefficient of determination, R²） - Lin氏一致性相关系数（Lin’s concordance correlation coefficient, CCC） - 均方预测误差（Mean squared prediction error, MSE） - 均方根预测误差（Root mean squared prediction error, RMSE） - 平均预测误差（偏差，Mean prediction error (bias)） rasters： 包含制图预测结果与相关不确定性的GeoTIFF栅格文件。这些文件按深度与输出类型进行组织。 深度约定与输出类型 预测生成自土壤表层至180 cm的标准深度区间。文件命名约定将这些区间记为d1至d10，对应如下深度切片： d1: 0–10 cm d2: 10–20 cm d3: 20–40 cm d4: 40–60 cm d5: 60–80 cm d6: 80–100 cm d7: 100–120 cm d8: 120–140 cm d9: 140–160 cm d10: 160–180 cm 每个深度切片包含三种预测类型： lower_percentile: 自助法预测的5%分位数（置信下限） 50th_percentile: 中位数（中心估计值） upper_percentile: 自助法预测的95%分位数（置信上限） 上述三类结果代表每个栅格单元内建模预测结果的不确定性区间。 参考文献：Malone B, Stockmann U, Glover M, McLachlan G, Engelhardt S, Tuomi S (2022). 数字土壤调查与制图：支撑土壤固有与动态属性状况评估. 土壤安全（Soil Security）, 6, 100048.