Non-visual common root rot disease detection using NIR spectrum and machine learning methods

Common root rot (CRR) is a soil-borne disease caused by Bipolaris sorokiniana in wheat contributing to significant yield losses in Australia. Detecting CRR is challenging due to its lack of visible symptoms above ground, necessitating time-consuming manual scouting. To address this issue, the potential of using non-destructive near-infrared (NIR) spectroscopy and machine learning models for early detection of CRR was explored. This study involved using a portable handheld NIR spectrometer to test five different wheat varieties with varying CRR resistance in the glasshouse and field trials. The machine learning methods of Logistic Regression (LR) and Support Vector Machines (SVM) with Principal Component Analysis (PCA) were compared with Deep Neural Networks (DNN) for the detection of CRR from NIR data. The results revealed that DNN outperformed LR and PCA-SVM models in classifying healthy and infected wheat plants, both in the glasshouse and field. The DNN achieved the highest classification accuracy, ranging from 68% to 85% in the glasshouse and reached the highest accuracy of 81% in the field at tillering stage. Moreover, spectral wavelengths in the range 1400-1700 nm with a focus on 1600-1700 nm were identified as highly indicative of the CRR. The combined use of NIR spectrometry and DNN demonstrated successful automated disease detection for CRR. These findings indicate the potential for a portable automated NIR sensing system for early crop disease detection, which could assist farmers in making informed management decisions regarding crop variety and fertilizer.