Enabling Precision Fertilisers Application Using Digital Soil Mapping in Australian Sugarcane Areas

Download files
Access & Terms of Use
open access
Copyright: Wang, Jie
Sugar is Australia's second largest export crop after wheat, generating a total annual revenue of almost $2 billion. It is produced from sugarcane, with approximately 95% grown in Queensland. While highly productive and contributing to the area’s economic sustainability, the soils in these areas have low fertility. The soils typically contain sand content > 60%, low organic carbon (SOC < 0.80%), cation exchange capacity (CEC), exchangeable Ca and Mg (< 8, 2.0, and 0.25 cmol(+) kg-1, respectively). Moreover, the soil is acidic (pH water < 5.5) and sodic (exchangeable sodium percentage [ESP] > 6%). Hence, sugarcane farmers need to apply fertilisers and ameliorants to maintain soil quality and productivity. Unfortunately, the high intensity rainfall in the region results in sediments, nutrients, and ameliorants run-off from these farms, resulting in environmental degradation and threats to marine ecology in the adjacent World Heritage Listed Great Barrier Reef. To mitigate these issues, the Australian sugarcane industry introduced the Six-Easy-Step Nutrient Management Guidelines. To apply these guidelines, a labour-intensive high-density soil sampling is typically required at the field level, followed by expensive laboratory analysis, spanning the myriad of biological, physical, and chemical properties of soils that need to be determined. To assist in sampling site selection, remote (e.g., Landsat-8, Sentinel-2, and DEM-based terrain attributes) and/or proximal sensing (e.g., electromagnetic [EM] induction and gamma-ray [γ-ray] spectrometry) digital data are increasingly being used. Moreover, the soil and digital data can be modelled using geostatistical (e.g., ordinary kriging [OK]), linear (e.g., linear mixed model [LMM]), machine learning (e.g., random forest [RF], quantile regression forest [QRF], support vector machine [SVM], and Cubist) and hybrid (e.g., RFRK, SVMRK, and CubistRK) approaches to enable prediction of soil properties from the rich source of digital data. However, there are many questions that need to be answered to determine appropriate recommendations including but not limited to i) which modelling approach is optimal, ii) which source of digital data is optimal and does fusion of various sources of digital data improve prediction accuracy, iii) which methods can be used to combine these digital data, iv) what is a minimum number of samples to establish a suitable calibration, v) which soil sampling designs could be used, and vi) what approaches are available to enable prediction of soil properties at various depths simultaneously? In this thesis, Chapter 1 introduces the research questions and defines the problems facing the Australian Sugarcane Industry in terms of the applications of the Six-Easy-Steps Nutrient Management Guidelines, research aims and thesis structure. Chapter 2 is a systematic literature review on various facets of DSM, which includes digital and soil data, models and outputs, and their application across various spatial scales and properties. In Chapter 3, prediction of topsoil (0-0.3 m) SOC is examined in the context of comparing predictive models (i.e., geostatistical, linear, machine learning [ML], and hybrid) using various digital data (i.e., remote [Landsat-8] and proximal sensors [EM and γ-ray]) either individually or in combination and determining minimum number of calibration samples. Chapter 4 shows to predict top- (0-0.3 m) and subsoil (0.6-0.9 m) Ca and Mg, various sampling designs (simple random [SRS], spatial coverage [SCS], feature space coverage [FSCS], and conditioned Latin hypercube sampling [cLHS]) were assessed, with different modelling approaches (i.e., OK, LMM, QRF, SVM, and CubistRK) and calibration sample size effect evaluated, using a combination of proximal data (EM and γ-ray) and terrain (e.g., elevation, slope, and aspect, etc.) attributes. Chapter 5 shows to enable the three-dimensional mapping of CEC and pH at topsoil (0-0.3 m), subsurface (0.3-0.6 m), shallow- (0.6-0.9 m) and deep-subsoil (0.9-1.2 m), an equal-area spline depth function can be used, with remote (Sentinel-2) and proximal data (EM and γ-ray) used alone or fused together, and various fusion methods (i.e., concatenation, simple averaging [SA], Bates-Granger averaging [BGA], Granger-Ramanathan averaging [GRA], and bias-corrected eigenvector averaging [BC-EA]) investigated. Chapter 6 explored the synergistic use of proximal (EM and γ-ray), and time-series of remote data (Landsat-8 and Sentinel-2) to map top- (0-0.15 m) and subsoil (0.30-0.45 m) ESP. The results show that, across these case studies, hybrid and ML models generally achieved higher prediction accuracy. The fusion of remote and proximal data produced better predictions, compared to single source of sensors. Granger-Ramanathan averaging (GRA) and concatenation were the most effective methods to combine digital data. A minimum of less than 1 sample ha-1 would be required to calibrate a good predictive model. There were differences in prediction accuracy amongst the sampling designs. The application of depth function splines enables the simultaneous mapping of soil properties from various depths. The produced DSM of soil properties can be used to inform farmers of spatial variability of soils and enable them to precisely apply fertilisers and/or ameliorants based on the Six-Easy-Step Nutrient Management Guidelines.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
Resource Type
Degree Type
PhD Doctorate
UNSW Faculty
download public version.pdf 35.8 MB Adobe Portable Document Format
Related dataset(s)