Please use this identifier to cite or link to this item: https://hdl.handle.net/1959.11/28167
Title: Embedded Machine-Learning For Variable-Rate Fertiliser Systems: A Model-Driven Approach To Precision Agriculture
Contributor(s): Stover, Joshua Marc  (author); Falzon, Gregory  (supervisor)orcid ; Lamb, David  (supervisor)
Conferred Date: 2019-10-02
Open Access: Yes
Handle Link: https://hdl.handle.net/1959.11/28167
Abstract: Efficient use of fertilisers, in particular the use of Nitrogen (N), is one of the rate-limiting factors in meeting global food production requirements. While N is a key driver in increasing crop yields, overuse can also lead to negative environmental and health impacts. It has been suggested that Variable-Rate Fertiliser (VRF) techniques may help to reduce excessive N applications. VRF seeks to spatially vary fertiliser input based on estimated crop requirements, however a major challenge in the operational deployment of VRF systems is the automated processing of large amounts of sensor data in real-time. Machine Learning (ML) algorithms have shown promise in their ability to process these large, high-velocity data streams, and to produce accurate predictions. The newly developed Fuzzy Boxes (FB) algorithm has been designed with VRF applications in mind, however no publicly available software implementation currently exists. Therefore, development of a prototype implementation of FB forms a component of this work. This thesis will also employ a Hardware-in-the-Loop (HWIL) testing methodology using a potential target device in order to simulate a real-world VRF deployment environment. By using this environment simulation, two existing ML algorithms (Artificial Neural Network (ANN) and Support Vector Machine (SVM)) can be compared against the prototype implementation of FB for applicability to VRF applications. It will be shown that all tested algorithms could potentially be suitable for high-speed VRF when measured on prediction time and various accuracy metrics. All algorithms achieved higher than 84.5% accuracy, with FB20 reaching 87.21%. Prediction times were highly varied; the fastest average predictor was an ANN (16.64μs), while the slowest was FB20(502.77μs). All average prediction times were fast enough to achieve a spatial resolution of 31 mm when operating at 60 m/s, making all tested algorithms fast enough predictors for VRF applications.
Publication Type: Thesis Masters Research
Field of Research (FoR): 070108 Sustainable Agricultural Development
079902 Fertilisers and Agrochemicals (incl. Application)
080399 Computer Software not elsewhere classified
Socio-Economic Objective (SEO): 890299 Computer Software and Services not elsewhere classified
HERDC Category Description: T1 Thesis - Masters Degree by Research
Appears in Collections:School of Science and Technology
Thesis Masters Research

Files in This Item:
6 files
File Description SizeFormat 
openpublished/StoverJoshuaMSc2019Thesis.pdfThesis1.63 MBAdobe PDF
Download Adobe
View/Open
Show full item record
Google Media

Google ScholarTM

Check


Items in Research UNE are protected by copyright, with all rights reserved, unless otherwise indicated.