Pre-visual common root rot disease detection in wheat using nir spectroscopy and uav-based multispectral imagery

Pre-visual common root rot disease detection in wheat using nir spectroscopy and uav-based multispectral imagery

This research presents innovative early detection methods for the soilborne disease common root rot (CRR) in wheat (Triticum aestivum L.) using advanced remote sensing technologies. Bipolaris sorokiniana-induced CRR is increasingly prevalent in wheat globally, reducing yields by 10–15% in susceptible cultivars due to indistinct above-ground symptoms. With current limited research, CRR may be more widespread in Australia than previously reported. This research addresses critical gaps by detailing CRR epidemiology, management strategies and potential detection methods. Currently, no effective control methods exist for CRR, making early disease detection beneficial for timely interventions to potentially mitigate yield losses. Remote sensing technologies involving Near-infrared (NIR) spectroscopy and Unmanned Aerial Vehicles (UAV)-based multispectral imagery are becoming promising for detecting diseases in large-scale fields due to their high accuracy, speed, nondestructive nature, and reduced labour compared to traditional scouting. While remote sensing technologies have been applied for detecting multiple visual wheat diseases at mid-to-late infection stages, early or pre-visual detection, especially for soilborne diseases remains underdeveloped. This thesis evaluated NIR spectroscopy and UAVbased multispectral imagery for CRR detection and severity classification to determine which method provided better accuracy, scalability, and ease of use as an alternative to traditional methods. Both technologies were combined with Deep Neural Networks (DNN) and advanced Machine Learning algorithms including Logistic Regression, Random Forest, Support Vector Machines and eXtreme Gradient Boosting. From 2021 to 2023, two glasshouse and three field trials over three agricultural cycles were conducted to differentiate between inoculated and non-inoculated wheat plants across five wheat varieties and to classify three disease severity levels. NIR spectroscopy using DNN algorithm developed from two seasons demonstrated optimal classification accuracy ranging from 66% to 91% in the glasshouse, and an average accuracy of 73% with up to 87% in the field. Important spectral wavelengths of 1600–1700 nm were identified as practical for detecting the presence of CRR, with initial detection occurring at 35 days after sowing in the glasshouse and 46 days in the field, both at the Z3 stem elongation stage before distinct visual symptoms appeared. UAV-based multispectral imagery analysis through Vegetation Indices (VIs) processed by pretrained DNN model distinguished CRR-inoculated and non-inoculated wheat plants with the highest accuracy of 93% at the Z3 stage and 79% overall accuracy. Pretrained DNN model from UAV-based multispectral imagery analysis classified three severity levels of CRR infection with 67% overall accuracy, which increased to 75% during the Z4-Z6 booting to anthesis stages. Important VIs for classification were chlorophyll- and RedEdge-based indices enabling the earliest detection at the Z3 stage and effective classification of disease severity at the Z4-Z6 stages. Further refined through Deep Transfer Learning, pre-trained DNN model from UAV-based multispectral imagery achieved 86.67% accuracy in disease detection for a different and unseen season. The results of two technologies in this research demonstrated that the Z3 stem elongation stage was the optimal earliest time for accurately detecting CRR in wheat. The use of NIR contact spectrometers and UAV-based multispectral cameras with DNN algorithms showed high potential in automated disease detection, which would assist in developing pre-visual symptom detection systems for CRR disease management on the farm. Future studies can emphasise algorithm refinement, advanced sensor technologies, and comparisons with abiotic stress factors to enhance the accuracy and applicability of CRR disease detection methods.

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