Accurate and detailed spatial soil information is essential for environmental modelling, risk assessment and decision making. The use of Remote Sensing data as secondary sources of information in digital soil mapping has been found to be cost effective and less time consuming compared to traditional soil mapping approaches. But the potentials of Remote Sensing data in improving knowledge of local scale soil information in West Africa have not been fully explored. This study investigated the use of high spatial resolution satellite data (RapidEye and Landsat), terrain/climatic data and laboratory analysed soil samples to map the spatial distribution of six soil properties–sand, silt, clay, cation exchange capacity (CEC), soil organic carbon (SOC) and nitrogen–in a 580 km2 agricultural watershed in south-western Burkina Faso. Four statistical prediction models–multiple linear regression (MLR), random forest regression (RFR), support vector machine (SVM), stochastic gradient boosting (SGB)–were tested and compared. Internal validation was conducted by cross validation while the predictions were validated against an independent set of soil samples considering the modelling area and an extrapolation area. Model performance statistics revealed that the machine learning techniques performed marginally better than the MLR, with the RFR providing in most cases the highest accuracy. The inability of MLR to handle non-linear relationships between dependent and independent variables was found to be a limitation in accurately predicting soil properties at unsampled locations. 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 in predicting soil properties. The results further showed that shortwave infrared and near infrared channels of Landsat8 as well as soil specific indices of redness, coloration and saturation were prominent predictors in digital soil mapping. Considering the increased availability of freely available Remote Sensing data (e.g. Landsat, SRTM, Sentinels), soil information at local and regional scales in data poor regions such as West Africa can be improved with relatively little financial and human resources.

No abstract

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.

No abstract

Abstract. The status of the soil organic carbon (SOC) stock at any position in the landscape is subject to a complex interplay of soil state factors operating at different scales and regulating multiple processes resulting either in soils acting as a net sink or net source of carbon. Forest landscapes are characterized by high spatial variability, and key drivers of SOC stock might be specific for sub-areas compared to those influencing the whole landscape. Consequently, separately calibrating models for sub-areas (local models) that collectively cover a target area can result in different prediction accuracy and SOC stock drivers compared to a single model (global model) that covers the whole area. The goal of this study was therefore to (1) assess how global and local models differ in predicting the humus layer, mineral soil, and total SOC stock in Swedish forests and (2) identify the key factors for SOC stock prediction and their scale of influence. We used the Swedish National Forest Soil Inventory (NFSI) database and a digital soil mapping approach to evaluate the prediction performance using random forest models calibrated locally for the northern, central, and southern Sweden (local models) and for the whole of Sweden (global model). Models were built by considering (1) only site characteristics which are recorded on the plot during the NFSI, (2) the group of covariates (remote sensing, historical land use data, etc.) and (3) both site characteristics and group of covariates consisting mostly of remote sensing data. Local models were generally more effective for predicting SOC stock after testing on independent validation data. Using the group of covariates together with NFSI data indicated that such covariates have limited predictive strength but that site-specific covariates from the NFSI showed better explanatory strength for SOC stocks. The most important covariates that influence the humus layer, mineral soil (0–50 cm), and total SOC stock were related to the site-characteristic covariates and include the soil moisture class, vegetation type, soil type, and soil texture. This study showed that local calibration has the potential to improve prediction accuracy, which will vary depending on the type of available covariates.