Using process-oriented model output to enhance machine learning-based soil organic carbon prediction in space and time

Monitoring and modelling soil organic carbon (SOC) in space and time can help us to better understand soil carbon dynamics and is of key importance to support climate change research and policy. Although machine learning (ML) has attracted a lot of attention in the digital soil mapping (DSM) community for its powerful ability to learn from data and predict soil properties, such as SOC, it is better at capturing soil spatial variation than soil temporal dynamics. By contrast, process-oriented (PO) models benefit from mechanistic knowledge to express physiochemical and biological processes that govern SOC temporal changes. Therefore, integrating PO and ML models seems a promising means to represent physically plausible SOC dynamics while retaining the spatial prediction accuracy of ML models. In this study, a hybrid modelling framework was developed and tested for predicting topsoil SOC stock in space and time for a regional cropland area located in eastern China. In essence, the hybrid model uses predictions of the PO model in unsampled years as additional training data of the ML model, with a weighting parameter assigned to balance the importance of SOC values from the PO model and real measurements. The results indicated that temporal trends of SOC stock modelled by PO and ML models were largely different, while they were notably similar between the PO and hybrid models. Cross-validation showed that the hybrid model had the best performance (RMSE = 0.29 kg m-2), with a 19 % improvement compared with the ML model. We conclude that the proposed hybrid framework not only enhances space-time soil carbon mapping in terms of prediction accuracy and physical plausibility, it also provides insights for soil management and policy decisions in the face of future climate change and intensified human activities.

Optimization of rain gauge sampling density for river discharge prediction using Bayesian calibration

River discharges are often predicted based on a calibrated rainfall-runoff model. The major sources of uncertainty, namely input, parameter and model structural uncertainty must all be taken into account to obtain realistic estimates of the accuracy of discharge predictions. Over the past years, Bayesian calibration has emerged as a suitable method for quantifying uncertainty in model parameters and model structure, where the latter is usually modelled by an additive or multiplicative stochastic term. Recently, much work has also been done to include input uncertainty in the Bayesian framework. However, the use of geostatistical methods for characterizing the prior distribution of the catchment rainfall is underexplored, particularly in combination with assessments of the influence of increasing or decreasing rain gauge network density on discharge prediction accuracy. In this article we integrate geostatistics and Bayesian calibration to analyze the effect of rain gauge density on river discharge prediction accuracy. We calibrated the HBV hydrological model while accounting for input, initial state, model parameter and model structural uncertainty, and also taking uncertainties in the discharge measurements into account. Results for the Thur basin in Switzerland showed that model parameter uncertainty was the main contributor to the joint posterior uncertainty. We also showed that a low rain gauge density is enough for the Bayesian calibration, and that increasing the number of rain gauges improved model prediction until reaching a density of one gauge per 340 km². While the optimal rain gauge density is case-study specific, we make recommendations on how to handle input uncertainty in Bayesian calibration for river discharge prediction and present the methodology that may be used to carry out such experiments.

Recent insights on uncertainties present in integrated catchment water quality modelling

This paper aims to stimulate discussion based on the experiences derived from the QUICS project (Quantifying Uncertainty in Integrated Catchment Studies). First it briefly discusses the current state of knowledge on uncertainties in sub-models of integrated catchment models and the existing frameworks for analysing uncertainty. Furthermore, it compares the relative approaches of both building and calibrating fully integrated models or linking separate sub-models. It also discusses the implications of model linkage on overall uncertainty and how to define an acceptable level of model complexity. This discussion includes, whether we should shift our attention from uncertainties due to linkage, when using linked models, to uncertainties in model structure by necessary simplification or by using more parameters. This discussion attempts to address the question as to whether there is an increase in uncertainty by linking these models or if a compensation effect could take place and that overall uncertainty in key water quality parameters actually decreases. Finally, challenges in the application of uncertainty analysis in integrated catchment water quality modelling, as encountered in this project, are discussed and recommendations for future research areas are highlighted.

Soil resources and element stocks in drylands to face global issues

Drylands (hyperarid, arid, semiarid, and dry subhumid ecosystems) cover almost half of Earth's land surface and are highly vulnerable to environmental pressures. Here we provide an inventory of soil properties including carbon (C), nitrogen (N), and phosphorus (P) stocks within the current boundaries of drylands, aimed at serving as a benchmark in the face of future challenges including increased population, food security, desertification, and climate change. Aridity limits plant production and results in poorly developed soils, with coarse texture, low C:N and C:P, scarce organic matter, and high vulnerability to erosion. Dryland soils store 646 Pg of organic C to 2 m, the equivalent of 32% of the global soil organic C pool. The magnitude of the historic loss of C from dryland soils due to human land use and cover change and their typically low C:N and C:P suggest high potential to build up soil organic matter, but coarse soil textures may limit protection and stabilization processes. Restoring, preserving, and increasing soil organic matter in drylands may help slow down rising levels of atmospheric carbon dioxide by sequestering C, and is strongly needed to enhance food security and reduce the risk of land degradation and desertification.

Soil nutrient maps of Sub-Saharan Africa: assessment of soil nutrient content at 250 m spatial resolution using machine learning

Spatial predictions of soil macro and micro-nutrient content across Sub-Saharan Africa at 250 m spatial resolution and for 0-30 cm depth interval are presented. Predictions were produced for 15 target nutrients: organic carbon (C) and total (organic) nitrogen (N), total phosphorus (P), and extractable-phosphorus (P), potassium (K), calcium (Ca), magnesium (Mg), sulfur (S), sodium (Na), iron (Fe), manganese (Mn), zinc (Zn), copper (Cu), aluminum (Al) and boron (B). Model training was performed using soil samples from ca. 59,000 locations (a compilation of soil samples from the AfSIS, EthioSIS, One Acre Fund, VitalSigns and legacy soil data) and an extensive stack of remote sensing covariates in addition to landform, lithologic and land cover maps. An ensemble model was then created for each nutrient from two machine learning algorithms- random forest and gradient boosting, as implemented in R packages ranger and xgboost-and then used to generate predictions in a fully-optimized computing system. Cross-validation revealed that apart from S, P and B, significant models can be produced for most targeted nutrients (R-square between 40-85%). Further comparison with OFRA field trial database shows that soil nutrients are indeed critical for agricultural development, with Mn, Zn, Al, B and Na, appearing as the most important nutrients for predicting crop yield. A limiting factor for mapping nutrients using the existing point data in Africa appears to be (1) the high spatial clustering of sampling locations, and (2) missing more detailed parent material/geological maps. Logical steps towards improving prediction accuracies include: further collection of input (training) point samples, further harmonization of measurement methods, addition of more detailed covariates specific to Africa, and implementation of a full spatiotemporal statistical modeling framework.

SoilGrids250m: Global gridded soil information based on machine learning

This paper describes the technical development and accuracy assessment of the most recent and improved version of the SoilGrids system at 250m resolution (June 2016 update). SoilGrids provides global predictions for standard numeric soil properties (organic carbon, bulk density, Cation Exchange Capacity (CEC), pH, soil texture fractions and coarse fragments) at seven standard depths (0, 5, 15, 30, 60, 100 and 200 cm), in addition to predictions of depth to bedrock and distribution of soil classes based on the World Reference Base (WRB) and USDA classification systems (ca. 280 raster layers in total). Predictions were based on ca. 150,000 soil profiles used for training and a stack of 158 remote sensing-based soil covariates (primarily derived from MODIS land products, SRTM DEM derivatives, climatic images and global landform and lithology maps), which were used to fit an ensemble of machine learning methods-random forest and gradient boosting and/or multinomial logistic regression-as implemented in the R packages ranger, xgboost, nnet and caret. The results of 10-fold cross-validation show that the ensemble models explain between 56% (coarse fragments) and 83% (pH) of variation with an overall average of 61%. Improvements in the relative accuracy considering the amount of variation explained, in comparison to the previous version of SoilGrids at 1 km spatial resolution, range from 60 to 230%. Improvements can be attributed to: (1) the use of machine learning instead of linear regression, (2) to considerable investments in preparing finer resolution covariate layers and (3) to insertion of additional soil profiles. Further development of SoilGrids could include refinement of methods to incorporate input uncertainties and derivation of posterior probability distributions (per pixel), and further automation of spatial modeling so that soil maps can be generated for potentially hundreds of soil variables. Another area of future research is the development of methods for multiscale merging of SoilGrids predictions with local and/or national gridded soil products (e.g. up to 50 m spatial resolution) so that increasingly more accurate, complete and consistent global soil information can be produced. SoilGrids are available under the Open Data Base License.

An integrated pan-tropical biomass map using multiple reference datasets

We combined two existing datasets of vegetation aboveground biomass (AGB) (Proceedings of the National Academy of Sciences of the United States of America, 108, 2011, 9899; Nature Climate Change, 2, 2012, 182) into a pan-tropical AGB map at 1-km resolution using an independent reference dataset of field observations and locally calibrated high-resolution biomass maps, harmonized and upscaled to 14 477 1-km AGB estimates. Our data fusion approach uses bias removal and weighted linear averaging that incorporates and spatializes the biomass patterns indicated by the reference data. The method was applied independently in areas (strata) with homogeneous error patterns of the input (Saatchi and Baccini) maps, which were estimated from the reference data and additional covariates. Based on the fused map, we estimated AGB stock for the tropics (23.4 N-23.4 S) of 375 Pg dry mass, 9-18% lower than the Saatchi and Baccini estimates. The fused map also showed differing spatial patterns of AGB over large areas, with higher AGB density in the dense forest areas in the Congo basin, Eastern Amazon and South-East Asia, and lower values in Central America and in most dry vegetation areas of Africa than either of the input maps. The validation exercise, based on 2118 estimates from the reference dataset not used in the fusion process, showed that the fused map had a RMSE 15-21% lower than that of the input maps and, most importantly, nearly unbiased estimates (mean bias 5 Mg dry mass ha(-1) vs. 21 and 28 Mg ha(-1) for the input maps). The fusion method can be applied at any scale including the policy-relevant national level, where it can provide improved biomass estimates by integrating existing regional biomass maps as input maps and additional, country-specific reference datasets.

Mapping Soil Properties of Africa at 250 m Resolution: Random Forests Significantly Improve Current Predictions

80% of arable land in Africa has low soil fertility and suffers from physical soil problems. Additionally, significant amounts of nutrients are lost every year due to unsustainable soil management practices. This is partially the result of insufficient use of soil management knowledge. To help bridge the soil information gap in Africa, the Africa Soil Information Service (AfSIS) project was established in 2008. Over the period 2008–2014, the AfSIS project compiled two point data sets: the Africa Soil Profiles (legacy) database and the AfSIS Sentinel Site database. These data sets contain over 28 thousand sampling locations and represent the most comprehensive soil sample data sets of the African continent to date. Utilizing these point data sets in combination with a large number of covariates, we have generated a series of spatial predictions of soil properties relevant to the agricultural management—organic carbon, pH, sand, silt and clay fractions, bulk density, cation-exchange capacity, total nitrogen, exchangeable acidity, Al content and exchangeable bases (Ca, K, Mg, Na). We specifically investigate differences between two predictive approaches: random forests and linear regression. Results of 5-fold cross-validation demonstrate that the random forests algorithm consistently outperforms the linear regression algorithm, with average decreases of 15–75% in Root Mean Squared Error (RMSE) across soil properties and depths. Fitting and running random forests models takes an order of magnitude more time and the modelling success is sensitive to artifacts in the input data, but as long as quality-controlled point data are provided, an increase in soil mapping accuracy can be expected. Results also indicate that globally predicted soil classes (USDA Soil Taxonomy, especially Alfisols and Mollisols) help improve continental scale soil property mapping, and are among the most important predictors. This indicates a promising potential for transferring pedological knowledge from data rich countries to countries with limited soil data.

Impact of input data uncertainty on environmental exposure assessment models: A case study for electromagnetic field modelling from mobile phone base stations

BACKGROUND

With the increased availability of spatial data and computing power, spatial prediction approaches have become a standard tool for exposure assessment in environmental epidemiology. However, such models are largely dependent on accurate input data. Uncertainties in the input data can therefore have a large effect on model predictions, but are rarely quantified.

METHODS

With Monte Carlo simulation we assessed the effect of input uncertainty on the prediction of radio-frequency electromagnetic fields (RF-EMF) from mobile phone base stations at 252 receptor sites in Amsterdam, The Netherlands. The impact on ranking and classification was determined by computing the Spearman correlations and weighted Cohen's Kappas (based on tertiles of the RF-EMF exposure distribution) between modelled values and RF-EMF measurements performed at the receptor sites.

RESULTS

The uncertainty in modelled RF-EMF levels was large with a median coefficient of variation of 1.5. Uncertainty in receptor site height, building damping and building height contributed most to model output uncertainty. For exposure ranking and classification, the heights of buildings and receptor sites were the most important sources of uncertainty, followed by building damping, antenna- and site location. Uncertainty in antenna power, tilt, height and direction had a smaller impact on model performance.

CONCLUSIONS

We quantified the effect of input data uncertainty on the prediction accuracy of an RF-EMF environmental exposure model, thereby identifying the most important sources of uncertainty and estimating the total uncertainty stemming from potential errors in the input data. This approach can be used to optimize the model and better interpret model output.

Incorporating DEM uncertainty in coastal inundation mapping

Coastal managers require reliable spatial data on the extent and timing of potential coastal inundation, particularly in a changing climate. Most sea level rise (SLR) vulnerability assessments are undertaken using the easily implemented bathtub approach, where areas adjacent to the sea and below a given elevation are mapped using a deterministic line dividing potentially inundated from dry areas. This method only requires elevation data usually in the form of a digital elevation model (DEM). However, inherent errors in the DEM and spatial analysis of the bathtub model propagate into the inundation mapping. The aim of this study was to assess the impacts of spatially variable and spatially correlated elevation errors in high-spatial resolution DEMs for mapping coastal inundation. Elevation errors were best modelled using regression-kriging. This geostatistical model takes the spatial correlation in elevation errors into account, which has a significant impact on analyses that include spatial interactions, such as inundation modelling. The spatial variability of elevation errors was partially explained by land cover and terrain variables. Elevation errors were simulated using sequential Gaussian simulation, a Monte Carlo probabilistic approach. 1,000 error simulations were added to the original DEM and reclassified using a hydrologically correct bathtub method. The probability of inundation to a scenario combining a 1 in 100 year storm event over a 1 m SLR was calculated by counting the proportion of times from the 1,000 simulations that a location was inundated. This probabilistic approach can be used in a risk-aversive decision making process by planning for scenarios with different probabilities of occurrence. For example, results showed that when considering a 1% probability exceedance, the inundated area was approximately 11% larger than mapped using the deterministic bathtub approach. The probabilistic approach provides visually intuitive maps that convey uncertainties inherent to spatial data and analysis.

SoilGrids1km--global soil information based on automated mapping

Background

Soils are widely recognized as a non-renewable natural resource and as biophysical carbon sinks. As such, there is a growing requirement for global soil information. Although several global soil information systems already exist, these tend to suffer from inconsistencies and limited spatial detail.

Methodology/Principal Findings

We present SoilGrids1km — a global 3D soil information system at 1 km resolution — containing spatial predictions for a selection of soil properties (at six standard depths): soil organic carbon (g kg−1), soil pH, sand, silt and clay fractions (%), bulk density (kg m−3), cation-exchange capacity (cmol+/kg), coarse fragments (%), soil organic carbon stock (t ha−1), depth to bedrock (cm), World Reference Base soil groups, and USDA Soil Taxonomy suborders. Our predictions are based on global spatial prediction models which we fitted, per soil variable, using a compilation of major international soil profile databases (ca. 110,000 soil profiles), and a selection of ca. 75 global environmental covariates representing soil forming factors. Results of regression modeling indicate that the most useful covariates for modeling soils at the global scale are climatic and biomass indices (based on MODIS images), lithology, and taxonomic mapping units derived from conventional soil survey (Harmonized World Soil Database). Prediction accuracies assessed using 5–fold cross-validation were between 23–51%.

Conclusions/Significance

SoilGrids1km provide an initial set of examples of soil spatial data for input into global models at a resolution and consistency not previously available. Some of the main limitations of the current version of SoilGrids1km are: (1) weak relationships between soil properties/classes and explanatory variables due to scale mismatches, (2) difficulty to obtain covariates that capture soil forming factors, (3) low sampling density and spatial clustering of soil profile locations. However, as the SoilGrids system is highly automated and flexible, increasingly accurate predictions can be generated as new input data become available. SoilGrids1km are available for download via http://soilgrids.org under a Creative Commons Non Commercial license.

Propagation of uncertainties in soil and pesticide properties to pesticide leaching

In the new Dutch decision tree for the evaluation of pesticide leaching to groundwater, spatially distributed soil data are used by the GeoPEARL model to calculate the 90th percentile of the spatial cumulative distribution function of the leaching concentration in the area of potential usage (SP90). Until now it was not known to what extent uncertainties in soil and pesticide properties propagate to spatially aggregated parameters like the SP90. A study was performed to quantify the uncertainties in soil and pesticide properties and to analyze their contribution to the uncertainty in SP90. First, uncertainties in the soil and pesticide properties were quantified. Next, a regular grid sample of points covering the whole of the agricultural area in the Netherlands was randomly selected. At the grid nodes, realizations from the probability distributions of the uncertain inputs were generated and used as input to a Monte Carlo uncertainty propagation analysis. The analysis showed that the uncertainty concerning the SP90 is 10 times smaller than the uncertainty about the leaching concentration at individual point locations. The parameters that contribute most to the uncertainty about the SP90 are, however, the same as the parameters that contribute most to uncertainty about the leaching concentration at individual point locations (e.g., the transformation half-life in soil and the coefficient of sorption on organic matter). Taking uncertainties in soil and pesticide properties into account further leads to a systematic increase of the predicted SP90. The important implication for pesticide regulation is that the leaching concentration is systematically underestimated when these uncertainties are ignored.

Using rainfall radar data to improve interpolated maps of dose rate in the Netherlands

The radiation monitoring network in the Netherlands is designed to detect and track increased radiation levels, dose rate more specifically, in 10-minute intervals. The network consists of 153 monitoring stations. Washout of radon progeny by rainfall is the most important cause of natural variations in dose rate. The increase in dose rate at a given time is a function of the amount of progeny decaying, which in turn is a balance between deposition of progeny by rainfall and radioactive decay. The increase in progeny is closely related to average rainfall intensity over the last 2.5h. We included decay of progeny by using weighted averaged rainfall intensity, where the weight decreases back in time. The decrease in weight is related to the half-life of radon progeny. In this paper we show for a rainstorm on the 20th of July 2007 that weighted averaged rainfall intensity estimated from rainfall radar images, collected every 5min, performs much better as a predictor of increases in dose rate than using the non-averaged rainfall intensity. In addition, we show through cross-validation that including weighted averaged rainfall intensity in an interpolated map using universal kriging (UK) does not necessarily lead to a more accurate map. This might be attributed to the high density of monitoring stations in comparison to the spatial extent of a typical rain event. Reducing the network density improved the accuracy of the map when universal kriging was used instead of ordinary kriging (no trend). Consequently, in a less dense network the positive influence of including a trend is likely to increase. Furthermore, we suspect that UK better reproduces the sharp boundaries present in rainfall maps, but that the lack of short-distance monitoring station pairs prevents cross-validation from revealing this effect.

Effect of land cover data on nitrous oxide inventory in fen meadows

Landscape representations based on land cover databases differ significantly from the real landscape. Using a land cover database with high uncertainty as input for emission inventory analyses can cause propagation of systematic and random errors. The objective of this study was to analyze how different land cover representations introduce systematic errors into the results of regional N2O emission inventories. Surface areas of grassland, ditches, and ditch banks were estimated for two polders in the Dutch fen meadow landscape using five land cover representations: four commonly used databases and a detailed field map, which most closely resembles the real landscape. These estimated surface areas were scaled up to the Dutch western fen meadow landscape. Based on the estimated surface areas agricultural N2O emissions were estimated using different inventory techniques. All four common databases overestimated the grassland area when compared to the field map. This caused a considerable overestimation of agricultural N2O emissions, ranging from 9% for more detailed databases to 11% for the coarsest database. The effect of poor land cover representation was larger for an inventory method based on a process model than for inventory methods based on simple emission factors. Although the effect of errors in land cover representations may be small compared to the effect of uncertainties in emission factors, these effects are systematic (i.e., cause bias) and do not cancel out by spatial upscaling. Moreover, bias in land cover representations can be quantified or reduced by careful selection of the land cover database.