Modelling soil physical and chemical properties using existing datasets across grain growing regions of Australia

Modelling soil physical and chemical properties using existing datasets across grain growing regions of Australia

Accurate characterisation and understanding of soil properties are crucial for evaluating soil fertility, crop water management, and sustaining crop production. Soil properties exhibit significant spatiotemporal variation due to the complex effects of natural and human factors. Major limitations to crop production in Australian soils include salinity, sodicity, acidity, alkalinity, and elemental toxicities (such as boron (B), chloride (Cl), and aluminium (Al)). Despite their importance, information on the extent and impact of these soil properties on Australian agriculture is often based on extrapolations from soil surveys and expert opinions specific to regions or is unavailable. There are several methods to assess soil properties: direct measurement, remote sensing, proximal soil sensing, and modelling. Direct measurement, involving soil sampling and laboratory analysis, is accurate but costly and time-consuming. Remote sensing, though useful, is limited to land surface observations and can be hindered by clouds and dense vegetation. Proximal sensing collects effective soil data but needs extensive lab-measured data for sensor calibration. Soil modelling combines data from various sources to understand soil systems. Modelling provides a useful method for estimating soil properties, though its accuracy often depends on data quality and may require recalibration for different regions or soil types. Additionally, many models are computationally demanding and rely on simplified assumptions, which can limit accessibility and affect reliability. This thesis explores the potential of a modelling approach using existing datasets to predict soil physical and chemical properties across Australian grain-growing regions. Firstly, this research develops a plant available water capacity (PAWC) prediction framework using Agricultural Production Systems sIMulator (APSIM), crop yield, farm management information and initial soil estimates. The framework involves configuring the APSIM model and optimising the lower limit of the soil profile to minimise the residual sum of squares between observed and predicted yields. A sensitivity test of the prediction framework against various input data sources was carried out. The result showed that the proposed PAWC prediction framework adequately predicts PAWC (R2 = 0.75, CCC = 0.75, RMSE = 47.95 mm, bias = -19.32 mm). When testing the framework’s sensitivity with different sources of climate and soil data, the results varied. Secondly, the cubist (CU) machine learning algorithm was applied to observed soil datasets to predict pH, Cl, electrical conductivity (EC), effective cation exchange capacity (ECEC) and exchangeable sodium percentage (ESP). The model demonstrated good accuracy for predicting soil pH, EC and ECEC, but faced challenges with Cl and ESP prediction. Thirdly, five machine learning models namely Random Forest (RF), CU, Extreme Gradient Boosting (XGB), Support Vector Machine (SVM) and K-Nearest Neighbour (kNN) are used to identify suitable models to predict Al, B, Cl and ESP. Extensive datasets including point-based soil data and raster-based environmental data, are used to model soil properties. Environmental covariates include climate variables such as annual precipitation and evaporation; parent material attributes like radiometric data, silica content, and weathering index; relief features derived from digital elevation models (DEM), including aspect, slope, and elevation; soil properties such as the Australian Soil Classification and mineral clays; and spectral information from barest Landsat imagery, including bands like blue, green, and SWIR1. Results revealed that the choice of machine learning model significantly impacts the uncertainty in predicting soil chemical properties. Specifically, RF algorithms exhibited the best accuracy in predicting these properties. The PAWC framework helps to predict PAWC, which is useful for making agronomic decisions but needs testing in constrained soils before future application. Machine learning models of soil chemical properties provide cost-effective alternatives to lab measurements, aiding in the identification of soil constraints. Accurate prediction of soil properties across diverse landscapes cost effectively promotes their widespread adoption. However, user-friendly interfaces or software tools should be developed in the future for practical application.

File reproduced in accordance with the copyright policy of the publisher/author/creator.