Pre-visual soilborne common root rot disease detection in wheat using UAV multispectral imagery and Deep Neural Networks

Common Root Rot (CRR), caused by Bipolaris sorokiniana , is a prevalent soilborne disease that severely impacts wheat production in Australia due to its difficult management. This study is the first to explore the potential of UAV-based multispectral imaging technologies for pre-visual soilborne CRR disease detection and severity classification in wheat across three seasons. Field multispectral imagery data were analysed using five algorithms: Logistic Regression (LR), Random Forest (RF), Support Vector Machines (SVM), eXtreme Gradient Boosting (XGBoost) and Deep Neural Networks (DNN). A Pre-trained DNN model developed from two seasons and validated in a third season achieved 93% accuracy in distinguishing CRR-inoculated wheat at the Z3 stem elongation stage and 79% overall accuracy. Moreover, the pre-trained DNN model classified three severity levels of CRR infection with 67% overall accuracy, which improved to 75% during Z4-Z6 (booting to anthesis) stages. Important Vegetation Indices for CRR disease detection and severity classification were chlorophyll- and RedEdge-based indices (PlantArea, Green, ExG, SCCCI and NDRE). The earliest CRR disease detection was achieved at the Z3 stage, with Z4-Z6 stages proving effective for severity classification. With further refinement, the pre-trained DNN model demonstrated effective validation for third-season disease detection, but not for severity classification. These findings could enable growers to improve field scouting by reducing reliance on labour-intensive manual scouting and subjective assessments, allowing them to adopt more effective disease management strategies and ultimately contributing to more sustainable wheat production. • ML models developed and validated via multi-season UAV multispectral wheat CRR disease trials. • DNN achieved 93% accuracy in distinguishing CRR-inoculated wheat as early as Z3 stem elongation. • DNN classified three CRR severities with 75% accuracy during Z4-Z6 booting to anthesis. • Chlorophyll -and RedEdge-based indices were key for CRR detection and severity classification.