Sie sind hier: Startseite Bonner Bodenk. Abh. Band 61 (2014) - Band 70 (2017) Band 69 (2017) Abstract Band 69

Abstract Band 69

Hounkpatin, O. (2017): Digital soil mapping using survey data and soil organic carbon dynamics in semi-arid Burkina Faso. 



With computer-assisted geostatistics and data mining methods, digital soil mapping (DSM) offers new possibilities for providing soil spatial information for data scarce areas such as West Africa. Such information could also be essential for understanding tropical soil organic carbon (SOC) sequestration potentials and dynamics. However, the level of accuracy depends on the statistical model selected, the choice of which is not clear from the first for such environments. Moreover, for datasets with imbalanced soil orders, prediction of reference soil groups (RSG) using a DSM approach often biased towards the majority soil order class. I hypothesized that (i) statistical models, which are able to handle both linear and unlinear patterns in data, will provide higher prediction accuracy than those geared towards linear patterns, (ii) pruning the major soil group - the Plinthosols - will result in increased prediction accuracy of the minor RSG, (iii) sites with savannah (SA) and related RSG will present larger SOC stocks than cropland (CR), however, (iv), with land use change (LUC) also the Plinthosols are prone to rapid SOC losses from bulk soil and primarily from coarse particle-size fractions.

To test these hypotheses, I sampled sites within both CR and SA across different RSG in the Dano catchment. For the DSM of soil properties (sand, silt, clay, CEC, SOC, N) in the topsoil (0 - 30 cm), four statistical prediction models – multiple linear regression (MLR), random forest regression (RF), support vector machine (SVM), stochastic gradient boosting (SGB) – were used and compared. To reduce the risk that the spatial prediction of the RSG was biased by the majority class – the Plinthosols – I used a data pruning approach, accounting for 80 %, 90 % and standard deviation core range of the Plinthosols data, respectively, while cutting off all data points belonging to the outer range. Random Forest was used as a robust data mining method along with its recursive feature elimination option to evaluate the performance of these different data subsets. The final assessment of SOC stocks was conducted by considering its variation in CR and SA and in various RSG at different depths. The spatial distribution of SOC stocks as well as the main related factors were then again elucidated using Random Forest. For understanding the temporal dynamics of SOC storage, I investigated a false chronosequence of Plinthosols that had been converted from SA to CR at a duration between 0 and 29 years. 

For the DSM of soil properties, results showed from the performance statistics that the machine learning techniques (RF, SVM, SGB) performed marginally better than the MLR, with the RF providing in most cases the highest accuracy. The lower performance of the MLR is attributed to its failure in accounting for non-linear relationships between response and predictor variables. The satellite data acquired during ploughing or early crop development stages (e.g. May, June) were found to be the most important spectral predictors, while elevation, temperature and precipitation came up as prominent terrain/climatic variables.


Upon the data pruning, the best predictions were observed when removing all PT points lower than 5 % and higher than 95  % of the cumulative percentage of the most important variable (wetness index). Modelling was then conducted solely with terrain and spectral parameters (TSP) with optimal predictors resulting from RF recursive feature elimination. The resulting prediction model provided a substantial agreement to observation, with a kappa value of 0.57 along with a 35  % increase in prediction accuracy for Cambisols, 16 % for Stagnosols and 7  % for Gleysols. The SAGA wetness index (S.Wet.Ind) was the most important variable driving the RSG suggesting that the humidity regime is a key discriminatory element among the RSG.  

The SOC stock distribution in the topsoil revealed a slightly larger SOC stock in the savannah sites (41.4 t C ha-1) than in the cropland (39.1 t C ha-1). Contrastingly, in the subsoil, a significant difference (p < 0.05) was observed between the CR recording a larger SOC stock of 40.2 t C ha-1, while the subsoil of the SA sites contained only 26.3 t C ha-1, on the average. Among the RSG, the Gleysols located at lower elevation positions revealed the largest SOC stocks over 0 - 30 cm (44 t C ha-1) and 100 cm depth (86.6 t C ha-1). Silt was the most abundant soil particle in the topsoil and was identified by the RF model as the most important factor related to the spatial distribution of the SOC stock, probably via its influence on soil moisture preservation and SOC storage via aggregation. Precipitation was found as the major factor related to subsoil SOC stock distribution. As the subsoils were also enriched in clay, the vertical transport of SOC rich sediments under tropical heavy rains likely accompanied major soil forming process in the landscape. 

The LUC in the chronosequence Plinthosols triggered losses in SOC stock of 24 t C ha-1 from the upper 10 cm and 49 t C ha-1 from the upper 30 cm. Thus, about 66 % (0 - 10 cm; p < 0.01) and 55 % (0 - 30 cm; p < 0.01) of the initial stock in the native vegetation had been released after 29 years of cultivation. Also, subsoil was found to be vulnerable to LUC, with SOC losses amounting on average to 0.7 to 19.5 t C ha-1 from the 30 - 100 cm depth interval. Losses of SOC occurred from all particle-size fractions with a mean residence time of SOC generally decreasing with increasing equivalent diameter of the particle-size fraction. In this study, I could not confirm Fe oxides as key factor influencing SOC stock stabilization, because only an average of 16 % of the total SOC stock were apparently bound to Fe.


In summary, DSM at local scale using RF with remote sensing data resulted in reasonable prediction accuracy for a large array of soil properties and RSG within a highly heterogeneous landscape. Data pruning proved to be efficient in a context where a RSG belonging to a wide range of terrain parameters overlapped with those related to only few RSG units. The SOC stocks as quantified in the present study reinforce the view that the semi-arid ecosystems of West Africa still offer an opportunity for carbon sequestration and these results represent a baseline for future modelling of SOC dynamics in the region. LUC from natural savannah to permanent cropland, however, affects both topsoil and subsoil SOC though the latter is scarcely considered in the impact analysis of LUC in Africa.