Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In a context where water availability is declining and competition for water is increasing, it is imperative to produce more crop per drop. To this end, rainfed and irrigated crop production systems must be designed which are adapted to the local soil, climate and socio-economic conditions. Integrated and well parameterized crop growth models are key to designing such crop production systems. Whereas today in most crop growth and hydrological modelling studies a lot of attention is devoted to collecting model inputs such as climate data, less attention is given to collecting and/or modelling of the crucial soil hydraulic properties (SHPs) data with sufficient spatial resolutions needed to operate the models over large areas. The collection of extensive data on the SHPs for territories as vast as the Zambezi River Basin (ZRB) is extremely labour intensive and costly, and so alternative approaches like Digital Soil Mapping (DSM) must be investigated. Digital soil mapping entails the generation of geographically referenced soil databases at a given resolution by using field observations and laboratory data methods coupled with environmental data. By applying geostatistical and/or machine learning techniques, quantitative relationships are identified between soil variables and environmental covariates to produce maps of soil properties and/or soil classes at different spatial resolutions. Therefore, relying on DSM, the general objective of this research project was to develop maps of soil hydraulic properties of the ZRB critical for crop growth and hydrological modelling.The research project was pursued in three major steps: (i) developing and evaluating pedotransfer functions (PTFs) for obtaining reference data of the soil hydraulic properties (SHP) water content at saturation (pF0.0), field capacity (pF2.0), and wilting point (pF4.2), besides available water content (AWC) and the saturated hydraulic conductivity (Ksat) for the ZRB, (ii) comparing DSM approaches to generate maps of SHPs for top soil and subsoil, covering the ZRB at 90 meter x 90 meter resolution and tapping a wide range of candidate covariates, and (iii) functionally evaluating the resulting maps by comparing the temporal evolution of the maize canopy cover simulated by means of the AquaCrop crop growth model alimented with the SHP-values retrieved from the digital soil maps with maize canopy cover derived from MODIS-LAI time series (2002 - 2012).As data of measured soil hydraulic properties (SHPs) are extremely scarce in the various soil databases covering the ZRB, including the Africa Soil Profiles database (AfSP), we first collected our own reference dataset through a sampling campaign in the Upper Mulungushi sub-basin (UMB) of the ZRB in Zambia. In the laboratory, we measured both basic soil characteristics such as soil granulometry and the SHPs. This data from 119 soil profiles, together with the data of 55 data points of the AfSP was used to develop pedotransfer functions for the SHPs (PTFs). The machine learning technique Artificial Neural Network (ANN)-PTF was found to yield the most accurate PTF.A functional evaluation of the PTFs with the FAO AquaCrop crop growth model, revealed that dry season irrigation requirements for maize computed with the ANN-PTFs were the closest to AquaCrop-outputs generated with measured soil hydraulic properties. Using the estimated SHPs from the ANN-PTFs as dependent variables, in the second step, we developed and compared different DSM models for spatially explicit estimation of hydraulic properties across the ZRB, and found that the best prediction method consisted of Random Forest as the deterministic model complemented with Residual Kriging (RK). Soil granulometry followed by climate and topographic elevation variables were the most important environmental covariates. In a final step, the digital maps of the SHP were also subjected to a functional evaluation using the FAO AquaCrop crop growth model. Canopy Cover (CC, m²m-2) derived from remotely sensed Leaf Area Index (LAI, m2m-2) from the MODIS (MCD15A3H version 6) archive was compared with (i) a time series of a maize CC simulated with AquaCrop crop and using SHPs estimated by with the digital soil maps, and (ii) by using SHP estimated by the Saxton and Rawls PTF. Pairwise comparison of the CC-time series - at 4 day temporal resolution - resulted in a RMSE of 0.07 m2m-2 and R2 of 0.93 for AquaCrop-CC-DSM versus MODIS-CC, and a RMSE of 0.08 m2m-2 and R2 of 0.88 for AquaCrop-CC-Saxton PTF versus MODIS-CC time series. Soil hydraulic properties estimated by the DSM maps and fed into the AquaCrop crop model, result in maize CC estimates that are closer to the MODIS CC than those based on SHP derived from the Saxton-PTFs.This research illustrates that there is still ample scope for further improving on both past and current mapping efforts. The wide availability of large sets of environmental data which can be used as covariates of soil properties, together with current powerful computational capacity allows to produce more detailed and more accurate maps than was done in the past. Though legacy soil data, such as available in the AfSP are invaluable for generating DSM - both at world, continental or regional scale - this research brings to light the importance of acquiring more field data for better predicting SHP and hence for developing more reliable crop growth and hydrological models. This research also shows that a functional evaluation of modelled data allows to better appreciate the performance of the model approach, allowing for a more complete judgement than merely assessing the models accuracy.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Following the increased economic importance of the sugarcane crop, a growing interest in monitoring growth and predicting yield of sugarcane has emerged. The time-consuming and pricey nature of collecting ground-based measurements and limitations in level of detail of satellite imagery make Unmanned Aerial Vehicles (UAVs) a potential interesting alternative remote sensing platform for the retrieval of biophysical crop parameters. UAVs can provide high spatial resolution imagery for a relatively low cost while having flexible deployment capabilities. The main objective of this research project was to evaluate the potential of RGB/NIR (Red-Blue-Green/Near-infrared) imagery acquired by UAVs as an alternative for ground observations. More specifically, four variables of interest – canopy height, leaf area index (LAI), stem density and biomass yield – were estimated by analysing UAV imagery in conjunction with ground truth data. Canopy height was estimated from a Digital Surface Model (DSM) generated from overlapping images taken by the UAV. The analysis showed that DSM derived canopy height and observed canopy height were highly correlated (R2 = 0.97). Furthermore, two decision tree-based regression techniques were calibrated and compared for their ability to predict LAI, stem density and biomass yield from the UAV imagery: random forest (RF) regression and boosted regression trees (BRT). The models were trained with a random subset representing 70% of the ground truth data, and then predictions were made on the remaining 30%. The predictive performance of various models was compared by calculating the Root Mean Squared Error (RMSE) and the R2 between the predicted and measured response variables. In general, the data analysis showed that despite imperfections in the data, these variables of interest can be estimated from the UAV images with a satisfactory accuracy. The most suitable model for LAI resulted in an RMSE of 0.72 (-) and R2 of 0.76, for stem density an RMSE of 1.43 stems/m and R2 of 0.92 and for biomass yield an RMSE of 7.35 ton/ha and R2 of 0.84.