PhD Thesis
Moisture scarcity is a limiting factor for sustainable agricultural productivity of dry
sub-humid agroecosystemsof sub-Saharan Africa (SSA). Designing sustainable
agricultural system management strategies responsive to the fluctuating soil
moisture regime is essential. Detailed and accurate information on soil moisture
storage conditions is essential for modelling agricultural system productivity.
Moisture storage capacity of the soils is quantified by moisture holding capacity
(MHC) which is defined as the difference between moisture content at field capacity
(FC)and wilting point (WP). Data availability is limited for SSA due to high costs
associated with direct measurement of MHC. Pedo-transfer functions (PTFs) and the
digital soil mapping (DSM)framework offer an opportunity for characterising spatial
variability of MHC through indirect approaches that integrate mathematical and
statistical methods. Though various methods exist for prediction and mapping MHC,
machine learning methods offer an avenue for more accurate characterisation of
MHC. The main objective of this study was to improve understanding on estimation
of soil moisture holding capacity at large spatial domains using machine learning
algorithms. This was achieved througha probabilistic sampling scheme,
development of MHC PTFs, and 3-dimensional characterisation of spatial variability
of MHC. One hundred (100) sampling locations were established over a geographic
area of about 44 km2by k-means clustering using R-statistical software. Two
sampling strategies were evaluated for optimisation of the sampling locations –a
stratified random sampling (STRS) and spatial coverage sampling (SPCS). Bulk soil
samples and soil cores were taken at three depth intervals of 0-30cm, 30-60 cm, and 60-100 cm at each sampling location. Geostatistical analysis and cross-validation
were performed for assessment of the sampling schemes using root mean square
error (RMSE), coefficient of determination (R2) and Mean Error (ME) as indices.
West-East anisotropy was evident in the MHC probably associated with topographic
and land cover effects. Spatial dependence ratio for the stratified random sampling
scheme (73 %) was higher than that of the spatial coverage sampling scheme (19 %).
This implied that SPCSdesign had better spatial correlation than the STRS design
due to a regular configuration of sampling nodes for SPCS design.Validation
resultswere better for STRS design than SPCS design. Pedo-transfer functions were
developed for FC and WP from support vector regression and multiple linear
regression with soil physico-chemical properties as predictors. Support vector
regression-PTFs had slightly better accuracy (RMSEs = 0.037 cm-3cm-3) than
multiple linear regression PTFs (RMSEs = 0.038 cm-3cm-3) and other published
PTFs. R2 values for SVR-PTFs were 66.3 and 67.9 % while those for MLR-PTFs
were 64.5 and 67.3% for FC and WP, respectively.Two machine learning algorithms
(Random forests(RF) and cubist decision trees (CB)) combined with soil depth
functions were evaluated for 3-dimensional mapping of MHC. Two DSM scenarios
were also evaluated (Measured data only (DSM-A) and measured plusPTF
estimated data (DSM-B)).Principal component analysis was performed on spatial
covariates layers representing soil forming factors for dimension reduction. Ten
principal components with a cumulative variance > 70 % were selected for mapping
process. Equal-area quadratic spline soil depth functions were fitted to model
continuous vertical distribution of MHC data. Prediction accuracy was good with
RMSEs ranging between 0.011-0.015 cm-3cm-3and R2 between 36 - 81.4 %. Random forests had better accuracy than the Cubist decision trees. A RF-CB ensemble
improves prediction accuracy.
Alliance for a Green Revolution in Africa (AGRA)