Digital soil mapping techniques across multiple landscape scales in South Africa
Thesis (PhDAgric)--Stellenbosch University, 2019.
ENGLISH ABSTRACT: Digital soil mapping has seen increasing interest due to environmental concerns and increasing food security issues. Digital soil mapping offers a quantitative approach which is cost effective as less soil observations are needed to produce large area soil maps. However, digital soil mapping has only recently been addressed in South Africa. This research aimed to produce two digital soil mapping (DSM) frameworks with the available resources in South Africa. The methodologies incorporate advanced geostatistics and/or machine learning techniques to be able to produce quantitative soil maps from the farm to catchment scale. First, a framework that optimises both feature selection and predictive models was developed to produce farm-scale soil property maps. Four feature selection techniques and eight predictive models were evaluated on their ability to predict particle size distribution and SOC. A boosted linear feature selection produced the highest accuracy for all but one soil property. The top performing predictive models were robust linear models for gravel (ridge regression, RMSE 9.01%, R2 0.75), sand (support vector machine, RMSE 4.69%, R2 0.67), clay (quantile regression, RMSE 2.38%, R2 0.52), and SOC (ridge regression, RMSE 0.19%, R2 0.41). Random forest was the best predictive model for silt content with a recursive feature selection (RMSE 4.12%, R2 0.53). This approach appears to be robust for farm-scale soil mapping where the number of observations is often small but high-resolution soil data is required. Second, 24 geomorphons (landform classification) were evaluated on their association with soil classes. The geomorphon with the highest association was aggregated into a 5-unit system which was evaluated on how well the system stratified soil lightness, soil EC, SOC, effective rooting depth, depth to lithology, gravel, sand, silt, and clay. It was found that an aggregated geomorphon stratified all soil attributes except EC. Additionally, the aggregated geomorphon predicted 6 out of 9 soil properties with the greatest accuracy (RMSE) when compared to the original geomorphon (10-unit system) and a manually delineated system (5-unit system). This study shows that aggregating geomorphons can stratify the soil landscape even at the farm-scale and can be used as an initial indication of the soil spatial variability. Third, a framework to disaggregate the Land Type Survey (LTS) through machine learning was developed. Geomorphons, together with the original LTS were overlaid to produce terrain morphological units. The polygons were disaggregated further to produce a raster map of soil depth classes through a disaggregation algorithm known as DSMART. The first most probable class raster achieved an accuracy of 68% and for the two most probable class rasters, an accuracy of 91% was achieved. The two-step approach proved necessary for producing a farm-scale soil map. Forth, a study aimed to compare 10 algorithms, implemented through a modified DSMART model, in their ability to disaggregate two polygons into soil associations in two environmentally contrasting locations (Cathedral Peak, KwaZulu-Natal Province and Ntabelanga, Eastern Cape Province). At Cathedral Peak (high relief with clear toposequences), nearest shrunken centroid was the top performing algorithm with a kappa of 0.42 and an average uncertainty of 0.22. At Ntabelanga (low relief with strong geological control), the results were unsatisfactory. However, a regularised multinomial regression was the top performing algorithm, achieving a kappa of 0.17 and an average uncertainty of 0.84. The results of this study highlight the versatility of a technique to disaggregate South Africa’s national resource inventory. Disaggregation was then used to simultaneously disaggregate 20 land types in the Mvoti catchment covering 317 km2 in KwaZulu Natal province. First, the optimal geomorphon was chosen through a spatially resampled Cramer’s V test to determine the association between the soil legacy polygons and the geomorphon units. Second, feature selection algorithms were embedded into DSMART. Third, the feature selection techniques were compared using 25, 50, 100, and 200 resamples per polygon. The results indicate that the Cramer’s V test is a rapid method to determine the optimal input map. Feature selection algorithms achieved the same accuracy as using all covariates but had greater computational efficiency. It is recommended that 10 to 20 times the amount of soil classes be used for the number of resamples per polygon.