NamSoil v1.0: Predicted Extractable Magnesium (mg kg-1) for Namibia at 90 m resolution (0–30, 30–60 and 60–100 cm)

Dataset Overview This dataset provides spatial predictions of Extractable Magnesium (mg kg-1) content across Namibia at 90 m spatial resolution for three standard soil depth intervals: 0–30, 30–60, 60–100 cm. For each depth interval, the following outputs are provided: predicted mean; 5th percentile; 95th percentile; 90% prediction interval (PI90). The maps are intended for national- and regional-scale applications and support environmental modelling, land evaluation, and resource management. Input Soil Data Model training was based on analytical data from the Namibian Soil Profile Database (NSPD2025) (https://zenodo.org/records/17618737). Profile locations have spatial accuracy better than 0.0001° and were reprojected to WGS84. Soil observations were depth-harmonised to the three standard depth intervals prior to modelling. Summary statistics of observed Extractable Magnesium (mg kg-1): 0–30 cm 30–60 cm 60–100 cm n 720 390 300 Min 0.00 0.00 0.00 Max 761.81 770.87 573.35 Mean 104.74 105.12 91.82 Median 77.97 66.68 47.23 SD 99.49 121.29 107.08 Skewness 2.20 2.13 1.71 Selected environmental covariates Environmental covariates included in the final model for each depth interval are: 0–30 cm: dem, tpi, chili, blue_w, green_w, swir1_w, swir2_w, ndvi_w, savi_w, msavi_w, evi_w, kndvi_w, blue_s, green_s, nir_s, savi_s, flow_lend_d, pet, arid_ind, aspp, aez, veg_types, aez_n, kaolinite, mafic, prec_wc2, tavg_wc2, geology_a, carb_diff, clay_diff, iron, rock_out 30–60 cm: chili, savi_w, msavi_w, evi_w, nir_s, evi_s, flow_lend_d, arid_ind, aez, cc, prec_wc2, tavg_wc2, geology_a, carb_diff, iron, rock_out 60–100 cm: dem, tpi, chili, blue_w, nir_w, evi_w, green_s, nir_s, evi_s, flow_lend_d, arid_ind, aspp, aez, prec_wc2, tavg_wc2, geology_a, carb_diff, iron, rock_out Full stack of environmental covariates Covariate Description dem Digital elevation model (altitude in metres) Slope Terrain gradient in degrees Aspect Slope facing direction (0–360°) Eastness East-west slope orientation (sin of aspect) Northness North-south slope orientation (cos of aspect) HorizontalCurvature Plan curvature; lateral flow convergence/divergence VerticalCurvature Profile curvature; flow acceleration along slope chili Continuous heat-insolation load index tpi Multi-scale topographic position index (ridges vs valleys) topo_diver Topographic diversity (habitat temperature/moisture variety) landforms_alos Hillslope position classes (15 landform types) flow_dir Local drainage flow direction hand Height above nearest drainage flow_accumul Upstream drainage area (km²) river_dist Distance to nearest drainage line flow_lend_d Flow length downstream to pour point flow_len_up Flow length upstream to farthest source landcover Land cover classes (11 classes, Sentinel-based) Prec_wc2 Mean annual precipitation 1970–2000 (mm) tavg_wc2 Mean annual temperature 1970–2000 (°C) arid_ind Aridity index (precipitation / potential evapotranspiration) pet Potential evapotranspiration 1970–2000 blue_s Landsat blue band (summer) blue_w Landsat blue band (winter) green_s Landsat green band (summer) green_w Landsat green band (winter) red_s Landsat red band (summer) red_w Landsat red band (winter) nir_s Landsat near-infrared band (summer) nir_w Landsat near-infrared band (winter) swir1_s Landsat shortwave infrared 1 (summer) swir1_w Landsat shortwave infrared 1 (winter) swir2_s Landsat shortwave infrared 2 (summer) swir2_w Landsat shortwave infrared 2 (winter) ndvi_s Normalized Difference Vegetation Index (summer) ndvi_w Normalized Difference Vegetation Index (winter) savi_s Soil Adjusted Vegetation Index (summer) savi_w Soil Adjusted Vegetation Index (winter) msavi_s Modified Soil Adjusted Vegetation Index (summer) msavi_w Modified Soil Adjusted Vegetation Index (winter) evi_s Enhanced Vegetation Index (summer) evi_w Enhanced Vegetation Index (winter) kndvi_s Kernel NDVI (summer) kndvi_w Kernel NDVI (winter) carb_diff Carbonate normalization ratio (Landsat) clay_diff Clay normalization ratio (Landsat) ferr_diff Ferrous minerals normalization ratio (Landsat) iron Iron normalization ratio (Landsat) rock_out Rock outcrop normalization ratio (Landsat) kaolinite index ASTER kaolinite mineral index calcite index ASTER calcite mineral index quartz index ASTER quartz mineral index carbonate index ASTER carbonate mineral index mafic index ASTER mafic mineral index Aez Agro-ecological zones of Namibia (1996, categorical) aez_n Updated agro-ecological zones of Namibia (2021) cc Potential carrying capacity of Namibia (2021) namsoil_13 National soil map (13 WRB reference soil groups) aspp Average seasonal plant productivity (1999–2019) veg_types Vegetation types geology_a Major rock groups by type and age geology Lithology units (geological map) Landform_iwa Iwahashi-Pike landform classification (slope, texture, convexity) convex Terrain convexity (ratio of positive curvature cells) curv_max Terrain curvature (rate of change in slope) The complete description and source details can be found in S5 – Environmental covariates assembled in the predictor stack.pdf file. Modelling Framework Spatial prediction was performed using the Random Forest algorithm. A bootstrap resampling strategy (20 iterations) was implemented, using an 80:20 calibration–validation split with replacement and a fixed random seed. Soil data preprocessing, hyperparameter tuning, feature selection, post-modelling metrics and external validation were executed in R, while covariate preparation, model implementation, and uncertainty quantification were conducted in Google Earth Engine. The Random Forest hyperparameters were: Depth interval ntree mtry nodesize sampsize 0–30 cm 150 10 6 0.56 30–60 cm 150 4 2 0.55 60–100 cm 150 1 3 0.73 where:ntree: number of decision trees in the forestmtry: the number of predictors randomly sampled at each RF splitnodesize: the minimum number of samples required at a leaf node to prevent overfittingsampsize: the in-bag (internal RF bootstrap) sample size drawn to train each tree Model Performance Model performance was evaluated for each bootstrap iteration using Root Mean Square Error (RMSE) to quantify prediction errors and Coefficient of Determination (R²) to measure explained variance. The performance metrics, averaged across the 20 bootstrap runs, are: Depth interval R² calibration RMSE calibration R² validation RMSE validation 0–30 cm 0.667 63.985 0.308 82.153 30–60 cm 0.814 63.359 0.268 101.761 60–100 cm 0.794 57.640 0.303 90.273 Uncertainty Quantification Uncertainty estimates were derived from the bootstrap prediction distributions. The 5th and 95th percentile maps represent lower and upper prediction limits. The 90% Prediction Interval Coverage Probability (PICP90) of Extractable Magnesium for the three depth classes were: Depth interval PICP90 0–30 cm 94.44 30–60 cm 93.85 60–100 cm 92.00 Data Outputs Map outputs are provided as Cloud-Optimised GeoTIFFs (WGS84) for GIS and modelling applications, and PNG format for visualisation and reporting. Data Access The input soil data used for model training is available in the Namibian Soil Profile Database (NSPD2025) at https://doi.org/10.5281/zenodo.17618737.Predicted soil maps can be retrieved directly from Zenodo using the quick-start scripts for reading, cropping, and exporting NamSoil layers — without downloading the full files — available at: https://github.com/Gelsleichter/acquire_NamSoil/.These scripts enable reproducible data retrieval workflows, allowing users to fetch and process specific layers programmatically. Code Availability The complete source code for data preprocessing, feature selection, hyperparameter tuning, model implementation, and post-processing is available at:https://doi.org/10.5281/zenodo.18776302, also published on https://github.com/Gelsleichter/NamSoil.The Google Earth Engine scripts for covariate preparation, regression matrix export, and Random Forest modelling with 20-iteration bootstrap are available at: https://code.earthengine.google.com/?accept_repo=users/Namibia_map/Soil_properties.Note that the GEE repository runs at a coarser spatial resolution than the published maps to reduce computational cost, memory usage, and export time within the Earth Engine environment. Users can adjust the output resolution to 90 m (or other) by modifying the scale parameter in the export functions, although this will require longer processing times and larger storage allocation.All scripts, fixed random seeds, and parameter configurations are provided to ensure full reproducibility of the modelling pipeline — from covariate preparation through spatial prediction and uncertainty quantification. Users can replicate the entire workflow or adapt individual components to other study areas or soil properties. Related Publication A full methodological description, model evaluation framework, and interpretation of results are provided in:[Publication DOI to be added]