iSDAsoil: soil sand content (USDA system) for Africa predicted at 30 m resolution at 0-20 and 20-50 cm depths

iSDAsoil dataset soil sand content in % predicted at 30 m resolution for 0–20 and 20–50 cm depth intervals. Data has been projected in WGS84 coordinate system and compiled as COG. Predictions have been generated using multi-scale Ensemble Machine Learning with 250 m (MODIS, PROBA-V, climatic variables and similar) and 30 m (DTM derivatives, Landsat, Sentinel-2 and similar) resolution covariates. For model training we use a pan-African compilations of soil samples and profiles (iSDA points, AfSPDB, LandPKS, and other national and regional soil datasets). Cite as: Hengl, T., Miller, M.A.E., Križan, J. et al. African soil properties and nutrients mapped at 30 m spatial resolution using two-scale ensemble machine learning. Sci Rep 11, 6130 (2021). https://doi.org/10.1038/s41598-021-85639-y To open the maps in QGIS and/or directly compute with them, please use the Cloud-Optimized GeoTIFF version. Layer description: sol_sand_tot_psa_m_30m_*..*cm_2001..2017_v0.13_wgs84.tif = predicted soil sand content mean value, sol_sand_tot_psa_md_30m_*..*cm_2001..2017_v0.13_wgs84.tif = predicted soil sand content model (prediction) errors, Model errors were derived using bootstrapping: md is derived as standard deviation of individual learners from 5-fold cross-validation (using spatial blocking). The model 5-fold cross-validation (mlr::makeStackedLearner) for this variable indicates: Variable: sand_tot_psa
R-square: 0.736
Fitted values sd: 22.8
RMSE: 13.7

Random forest model:
Call:
stats::lm(formula = f, data = d)

Residuals:
Min 1Q Median 3Q Max
-80.626 -5.321 0.221 6.071 88.686

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 6.687471 24.022001 0.278 0.780714
regr.ranger 1.060521 0.003503 302.742 < 2e-16 ***
regr.xgboost -0.018718 0.004910 -3.812 0.000138 ***
regr.cubist 0.031749 0.003922 8.096 5.73e-16 ***
regr.nnet -0.161127 0.422746 -0.381 0.703098
regr.cvglmnet -0.028217 0.004462 -6.323 2.57e-10 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 13.65 on 122261 degrees of freedom
Multiple R-squared: 0.736, Adjusted R-squared: 0.736
F-statistic: 6.818e+04 on 5 and 122261 DF, p-value: < 2.2e-16 To submit an issue or request support please visit https://isda-africa.com/isdasoil

iSDAsoil数据集针对0–20cm与20–50cm两个深度分层，以30米分辨率预测得到土壤砂粒含量（百分比形式）。数据采用WGS84坐标系进行投影，并整理为云优化GeoTIFF（Cloud-Optimized GeoTIFF，COG）格式。

本数据集的预测结果基于多尺度集成机器学习生成，协变量包含两类分辨率特征：250米分辨率数据（如MODIS、PROBA-V、气候变量等）以及30米分辨率数据（如数字地形模型衍生数据、Landsat、Sentinel-2等）。

模型训练采用泛非地区的土壤样本与剖面数据集，包括iSDA采样点、AfSPDB、LandPKS及其他国家级、区域级土壤数据集。

### 引用格式
Hengl, T., Miller, M.A.E., Križan, J. 等. 采用双尺度集成机器学习方法绘制30米空间分辨率的非洲土壤属性与养分分布图. 科学报告, 11, 6130 (2021). https://doi.org/10.1038/s41598-021-85639-y

### 图层使用说明
若需在QGIS中打开该地图数据或直接进行计算，请使用云优化GeoTIFF版本。各图层含义如下：
- `sol_sand_tot_psa_m_30m_*..*cm_2001..2017_v0.13_wgs84.tif`：预测土壤砂粒含量的均值
- `sol_sand_tot_psa_md_30m_*..*cm_2001..2017_v0.13_wgs84.tif`：土壤砂粒含量预测模型的误差。模型误差通过自助法（bootstrapping）推导得到：`md`为5折空间阻塞交叉验证中各基学习器的标准差。

该变量的5折交叉验证结果（基于`mlr::makeStackedLearner`实现）如下：
- 变量：sand_tot_psa
- 决定系数（R-square）：0.736
- 拟合值标准差：22.8
- 均方根误差（RMSE）：13.7

#### 随机森林集成元模型回归结果
调用形式：`stats::lm(formula = f, data = d)`

残差统计量：
最小值 第一四分位数 中位数 第三四分位数 最大值
-80.626 -5.321 0.221 6.071 88.686

系数表：
变量名 估计值 标准误差 t值 P值
(Intercept) 6.687471 24.022001 0.278 0.780714
regr.ranger 1.060521 0.003503 302.742 < 2e-16 ***
regr.xgboost -0.018718 0.004910 -3.812 0.000138 ***
regr.cubist 0.031749 0.003922 8.096 5.73e-16 ***
regr.nnet -0.161127 0.422746 -0.381 0.703098
regr.cvglmnet -0.028217 0.004462 -6.323 2.57e-10 ***

---
显著性代码：0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

残差标准误差：13.65，自由度为122261
多重决定系数（Multiple R-squared）：0.736，调整后决定系数（Adjusted R-squared）：0.736
F统计量：6.818×10^4，自由度为5和122261，P值 < 2.2×10^-16

若需提交问题或寻求技术支持，请访问 https://isda-africa.com/isdasoil