Uncertainty-guided sampling to improve digital soil maps

Stumpf, Felix; Schmidt, Karsten; Goebes, Philipp; Behrens, Thorsten; Schönbrodt-Stitt, Sarah; Wadoux, Alexandre M.J.‐C.; Xiang, Wei; Scholten, Thomas

doi:10.1016/j.catena.2017.01.033

Cited by 39 publications

(25 citation statements)

References 68 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We applied random forest (RF) regression, an ensemble classifier that is based on averaging the results of multiple randomized decision tree models for the final estimations (Peters et al, 2007;Breiman, 2001). RF was selected because it includes an internal error estimation and has been successfully applied in the field of DSM (Peters et al, 2007;Heung et al, 2014;Stumpf et al, 2015b). We set-up RF regression models for each required steady soil parameter using the 140 topsoil samples and a pool of continuous terrain covariates, because RF is robust to noise and multi-collinearity in the predictors (Díaz-Uriarte & De Andrés, 2006).…”

Section: Modelling Sediment Reallocationmentioning

confidence: 99%

Sediment Reallocations due to Erosive Rainfall Events in the Three Gorges Reservoir Area, Central China

et al. 2016

Self Cite

View full text Add to dashboard Cite

Soil erosion by water outlines a major threat to the Three Gorges Reservoir Area in China. A detailed assessment of soil conservation measures requires a tool that spatially identifies sediment reallocations due to rainfall-runoff events in catchments. We applied EROSION 3D as a physically based soil erosion and deposition model in a small mountainous catchment. Generally, we aim to provide a methodological frame that facilitates the model parametrization in a data scarce environment and to identify sediment sources and deposits. We used digital soil mapping techniques to generate spatially distributed soil property information for parametrization. For model calibration and validation, we continuously monitored the catchment on rainfall, runoff and sediment yield for a period of 12 months. The model performed well for large events (sediment yield > 1 Mg) with an averaged individual model error of 7.5%, while small events showed an average error of 36.2%. We focused on the large events to evaluate reallocation patterns. Erosion occurred in 11.1% of the study area with an average erosion rate of 49.9 Mg ha À1 . Erosion mainly occurred on crop rotation areas with a spatial proportion of 69.2% for 'corn-rapeseed' and 69.1% for 'potato-cabbage'. Deposition occurred on 11.0%. Forested areas (9.7%), infrastructure (41.0%), cropland (corn-rapeseed: 13.6%, potatocabbage: 11.3%) and grassland (18.4%) were affected by deposition. Because the vast majority of annual sediment yields (80.3%) were associated to a few large erosive events, the modelling approach provides a useful tool to spatially assess soil erosion control and conservation measures.

show abstract

Section: Modelling Sediment Reallocationmentioning

confidence: 99%

Sediment Reallocations due to Erosive Rainfall Events in the Three Gorges Reservoir Area, Central China

et al. 2016

Self Cite

View full text Add to dashboard Cite

show abstract

“…As a result, there is a potential risk that a bias may be introduced when discriminating between soil classes in the field 50,51 . These errors are mostly formulated as source errors 52 .…”

Section: Introductionmentioning

confidence: 99%

Predicting reference soil groups using legacy data: A data pruning and Random Forest approach for tropical environment (Dano catchment, Burkina Faso)

Hounkpatin

Schmidt

Stumpf

et al. 2018

Sci Rep

Self Cite

View full text Add to dashboard Cite

Predicting taxonomic classes can be challenging with dataset subject to substantial irregularities due to the involvement of many surveyors. A data pruning approach was used in the present study to reduce such source errors by exploring whether different data pruning methods, which result in different subsets of a major reference soil groups (RSG) – the Plinthosols – would lead to an increase in prediction accuracy of the minor soil groups by using Random Forest (RF). This method was compared to the random oversampling approach. Four datasets were used, including the entire dataset and the pruned dataset, which consisted of 80% and 90% respectively, and standard deviation core range of the Plinthosols data while cutting off all data points belonging to the outer range. The best prediction was achieved when RF was used with recursive feature elimination along with the non-oversampled 90% core range dataset. This model provided a substantial agreement to observation, with a kappa value of 0.57 along with 7% to 35% increase in prediction accuracy for smaller RSG. The reference soil groups in the Dano catchment appeared to be mainly influenced by the wetness index, a proxy for soil moisture distribution.

show abstract

“…()) would serve in‐depth comparison between assessments. Several methods exist to estimate uncertainty, for example through stochastic Monte‐Carlo simulations (Chartin et al., ) or decision trees (Stumpf et al., ). Such measures can be used to guide and optimize soil sampling (Stumpf et al., ).…”

Section: Discussionmentioning

confidence: 99%

The devil is in the detail: Discrepancy between soil organic carbon stocks estimated from regional and local data sources in Flanders, Belgium

Ottoy

Vanierschot

Dondeyne

et al. 2019

Soil Use and Management

View full text Add to dashboard Cite

Storage of soil organic carbon (SOC) is an essential function of ecosystems underpinning the delivery of multiple services to society. Regional SOC stock estimates often rely on data collected during land‐use‐specific inventory schemes with varying sampling depth and density. Using such data requires techniques that can deal with the associated heterogeneity. As the resulting SOC assessments are not calibrated for the local scale, they could suffer from oversimplification of landscape processes and heterogeneity. This might especially be the case for sandy regions where typical historical land use practices and soil development processes determine SOC storage. The aims of this study were (a) to combine four land‐use‐specific SOC stock assessments to estimate the total stock in Flanders, Belgium, and (b) to evaluate the applicability of this regional‐scale estimate at the local scale. We estimated the SOC stock in the upper 100 cm of the unsealed area in Flanders (887,745 ha) to be 111.67 Mt OC, or 12.6 ± 5.65 kg OC m−2 on average. In general, soils under (semi‐) natural land‐use types, for example forests, store on average more organic carbon than under agriculture. However, overall agricultural soils store the largest amounts of SOC due to their vast spatial extent. Zooming in on a sandy location study (13.55 km2) revealed the poor performance of the regional estimates, especially where Histosols occurred. Our findings show that a greater spatial sampling density is required when SOC stock estimates are needed to inform carbon‐aware land management rather than to provide for regional reporting.

show abstract

Uncertainty-guided sampling to improve digital soil maps

Cited by 39 publications

References 68 publications

Sediment Reallocations due to Erosive Rainfall Events in the Three Gorges Reservoir Area, Central China

Sediment Reallocations due to Erosive Rainfall Events in the Three Gorges Reservoir Area, Central China

Predicting reference soil groups using legacy data: A data pruning and Random Forest approach for tropical environment (Dano catchment, Burkina Faso)

The devil is in the detail: Discrepancy between soil organic carbon stocks estimated from regional and local data sources in Flanders, Belgium

Contact Info

Product

Resources

About