Discrimination of wheat crown rot utilising wavelet based models in the NIR spectrum

Discrimination of wheat crown rot utilising wavelet based models in the NIR spectrum

Crown rot is a stubble-borne disease of significant economic importance to the wheat industry, with no rapid detection method and human scoring only possible late in disease development after visual symptoms appear. This research aimed to develop non-destructive plant sensing tools to improve crown rot screening and accelerate development of resistant and tolerant germplasm. Machine learning models were developed for the discrimination and quantification of Fusarium pseudograminearum induced crown rot from contact near-infrared sensor data (900–1700 nm) in three glasshouse and two field trials, and near-infrared camera data (900–1700 nm) in one glasshouse trial. Contact sensor data was modeled with principal component analysis (PCA) and the discrete wavelet transform (DWT), and the impact of sensing location (i.e. plant part), timing of sensing and different training splits were compared. Specific models were generated by grouping sensor data into weekly intervals for each trial. Generalised models were generated by combining data across multiple trial sites and temporal windows (weekly, three-weekly and whole-season).DWT achieved higher crown rot detection accuracy than PCA in 67% of test cases for contact sensing when training on 20% of the disease data for specific models, with maximum average accuracies of 70.5%. Both DWT and PCA performed best in 50% of test cases when trained on the 80-20% split, with maximum accuracies of 75.7% for DWT and 76.9% for PCA. PCA was more accurate than DWT in a majority of individual test cases across both data splits for generalised models, with maximum accuracies of 69.8%. PCA utilised fewer overall features successfully in both specific and generalised models on both training splits. Significant differences in accuracy between sensing dates, sensing locations and interaction were determined. The differences were examined using DWT and six machine learning methods for generalised models, with an 80-20% train-test split. Individual readings and combinations of tiller, centre and flag leaf were evaluated. The highest performing combination of measurements achieved an average of 77.9% accuracy across all five trials using the contact sensor. Phenotyping capability was examined using a multilayer perceptron preprocessed with DWT and PCA to develop models to quantify disease severity. Six severity scales based on human scoring of visual symptoms were developed and tested on contact sensor data from glasshouse and field trials. Both PCA and DWT performed similarly with individual models obtaining mean model accuracy ranging from 32% to 96%. Models performed best in field environments. Near-infrared image data was collected in a glasshouse trial across four weeks of early crown rot infection, using narrow bandpass filters centered at five wavebands identified using the contact sensor. Single and multi-input convolutional neural networks were created for discrimination and quantification of crown rot infection. The discrimination model achieved average accuracies of 53–100%, with highest average accuracies obtained in weeks 2–4. The quantification model achieved average accuracies of 73% when trained on combined data across all weeks. Developed models were successfully ported onto a mobile development board for real-time detection applications. It is concluded that successful detection and quantification of crown rot was achieved using both contact and camera-based near-infrared sensing. These findings are the initial steps in developing a high-throughput phenotyping system to provide wheat breeders with new tools and methods for crown rot resistance breeding. Further work should evaluate developed models on a wide range of germplasm and extend to real-time model execution.