BIS-4D: Maps of soil properties and their uncertainties at 25 m resolution in the Netherlands

Samenvatting

This dataset is an asset of the scientific manuscript "BIS-4D: Mapping soil properties and their uncertainties at 25m resolution in the Netherlands" (Helfenstein et al., 2024, under review). It contains maps of soil properties and their uncertainties at 25m resolution in the Netherlands obtained using the BIS-4D soil modelling and mapping platform. BIS-4D is based on well-established digital soil mapping practices. This dataset includes maps of predictions of the mean, 0.05, 0.50 (median) and 0.95 quantiles and the 90th prediction interval width (PI90) of clay content [%], silt content [%], sand content [%], bulk density (BD) [g/cm3], soil organic matter (SOM) [%], pH [KCl], total N (Ntot) [mg/kg], oxalate-extractable P (Pox) [mmol/kg] and cation exchange capacity (CEC) [mmol(c)/kg]. Prediction maps are available for the standard depth layers specified by the GlobalSoilMap initiative (0-5, 5-15, 15-30, 30-60, 60-100 and 100-200cm). For SOM, these prediction maps are available for the years 1953, 1960, 1970, 1980, 1990, 2000, 2010, 2020 and 2023 based on changing land use, peat classes and peat occurrence over time. BIS-4D uses georeferenced soil point data (field estimates and laboratory measurements), spatially explicit environmental variables (covariates), and machine learning to predict in 3D space, and for SOM, in 3D space and time. More information about how these maps were created, the BIS-4D soil modelling and mapping platform, accuracy assessment, strengths, limitations, map assessment scale and specific user recommendations can be found in the scientific paper "BIS-4D: Mapping soil properties and their uncertainties at 25m resolution in the Netherlands" (Helfenstein et al., 2024, under review). The BIS-4D model code is available on GitLab. Please note that an earlier version of soil pH prediction maps were published. In comparison, this version contains several important updates. Firstly, covariates of peat classes, groundwater classes in agricultural areas and Sentinel 2 RGB and NIR bands and spectral indices were added, all of which were selected and thus used for model calibration and prediction of the updated BIS-4D prediction maps. We also included de-correlation and recursive feature elimination to increase the signal to noise ratio, make models more parsimonious and increase reproducibility. Please consider the following file naming structure to make it easier to find the prediction maps you need: File naming structure: "[soil property]_d_[upper depth layer boundary]_[lower depth layer boundary]_QRF_[PI90/pred type]_[processed].tif" Example: "clay_per_d_0_5_QRF_pred_mean_processed.tif" Soil property denotes the target soil property (listed above), depth upper and lower boundaries indicate the prediction target depth, QRF = quantile regression forest, which is the algorithm used for model calibration and prediction, PI90 is a measure of prediction uncertainy and is the 95th - 5th quantile, "pred_mean" indicates mean predictions, "pred50" indicates median predictions, "pred5" indicates 5th quantile prediction and "pred95" indicates 95th quantile prediction. For clay, silt and sand content, predictions were post-processed so that they add up to 100% and therefore for those GeoTIFF files the names contain "_processed". For SOM, the target prediction year is also indicated directly after "SOM_per", e.g. "SOM_per_2023_d_0_5_QRF_pred_mean.tif".