Detection of crown rot in wheat utilising near-infrared spectroscopy: towards remote and robotic sensing

Abstract

Forty percent wheat yield reduction is reported globally due to crown rot (Fusarium pseudograminearum). An emerging approach for sensor-based disease discrimination is the use of spectral reflectance with combinations of wavebands and varying bandwidths, which has potential to reduce the impact of environmental factors on spectral sensitivity detection accuracy. Transferring such technology from a laboratory to field environment presents challenges, particularly in regard to producing adequately robust models.An experiment was conducted in which near-infrared spectral reflectance data was captured in a glasshouse environment, for cultivars of winter wheat with varying resistances to F. pseudograminearum. A contact sensor sensitive to shortwave infrared (900-1700 nm) wavebands was used. Raw sensor data was calibrated and transformed using a discrete wavelet approach, allowing for variable waveband size. Optimised machine learning disease identification models were compared across the nine weeks following inoculation with F. pseudograminearum. Models were compared for the ability to accurately detect crown rot across weeks.The results show crown rot detection ability with accuracies ranging from 49-74%, as well as a temporal patterning effect as the season progresses. An artificial neural network classifier (ANN) performed best with a top accuracy of 74.14%, of the six machine learning algorithms trialed. Waveform differences between plus and minus treatments indicate that the sensing approach has potential to be scaled to a camera-based system for use on remote sensing platforms. Further work is being conducted to understand the viability of such an approach, which is an important step towards both robotic and RPA-based disease discrimination.

The Promise of Hyperspectral Imaging for the Early Detection of Crown Rot in Wheat

“…ANN can analyze and process image data through a neural network with input vectors and output vectors (neurons/nodes) [98]. The application ANN distinguished healthy versus infected wheat plant in near-infrared (900-1700 nm) images and obtained an accuracy of 74.14%, after nine weeks of infection [99]. Such several studies indicate the crown rot could be detected by the HSI as summarized in Table 1, and there is high probability to develop a rapid method to detection crown by HSI at infection early stage.…”

Proximal Methods for Plant Stress Detection Using Optical Sensors and Machine Learning

“…ANNs have been used successfully in many studies for the identification and classification of various plant stresses. These include detecting powdery mildew and soft rot in zucchini [129], classifying biotic stresses in pomegranate [106], detecting orange spotting disease in oil palm [130], and identifying crown rot in wheat [131]. A major advantage of ANNs is their ability to be used without specialized knowledge on the data and its interpretation; however, disadvantages include being prone to overfitting and requiring greater amounts of computational resources [132].…”

Remote detection of Fusarium crown rot in broadacre bread wheat and durum wheat through use of aerial imagery

“…These changes are often observed in the visible spectrum (VIS) as well as red edge and near-infrared (NIR) regions (380-780 nm, 650-780 nm and 800-2500 nm, respectively) (Franke and Menz 2007). NIR using five bands (900-1700 nm) has been demonstrated for non-destructive discrimination between Fp-infected and non-infected wheat plants at the seedling stage with an accuracy of 55-100% at 3-11 weeks after infection under glasshouse conditions (Humpal et al 2020a). Changes in spectral reflectance are only indicative of a change in homogeneity of crop health, size or biomass and are unable to confirm infection or cause of disease without groundtruthing of locations of interest (Bhandari et al 2018;Nagai et al 2020).…”

“…This study investigated the potential of both multispectral and thermal (8-14 μm) reflectance to detect FCR in bread wheat (Triticum aestivum L.) and durum wheat (T. durum Desf.). Previous work has been conducted to explore the prospects for FCR detection in wheat using NIR sensors (Humpal et al 2020b). However, the success of that study was limited to controlled environments with contact NIR sensors, because the accuracy decreased significantly when the sensors were removed from the plant tissue.…”