Rice bacterial leaf blight detection using canopy hyperspectral data with spectral transformation methods

Rice, the staple food for over half of the global population, is vital for food security and regional economies, especially in developing countries. Bacterial leaf blight (BLB), caused by Xanthomonas oryzae pv. oryzae (Xoo), significantly reduces yields under favourable weather conditions. Timely detection of BLB is crucial for effective management. Remote sensing, especially hyperspectral measurements offers a rapid and non-destructive alternative to identify stress-induced variations in leaves and canopies. With hundreds of narrow spectral bands, hyperspectral data can assist in distinguishing between healthy and BLB-infected rice based on significant differences in the rice spectral signatures. This study evaluated canopy hyperspectral measurements obtained from inoculated rice fields for detecting BLB. Besides the original spectra, spectral transformation methods including first derivative, second derivative, continuum removal (CR) and continuous wavelet transform (CWT) were tested to distinguish healthy and BLB-infected rice. The sensitivity of spectral bands to BLB was examined on both original and transformed spectra. The most sensitive spectral bands were identified using the optimal thresholding method, with the top 5% most accurate bands highlighted. Results showed that the most sensitive spectral bands are mainly identified from original and continuum-removed spectra compared with other spectral transformation methods. Using the original spectra, sensitive bands were concentrated in the red and red edge regions. While for continuum-removed CR spectra, sensitive bands were also concentrated in the near-infrared region. The highest classification accuracy was achieved by continuum-removed spectra at 970 nm for differentiating healthy and BLB-infected rice (Overall accuracy=79%). These findings demonstrate the potential of CR for enhanced diagnosis of BLB in rice. Future efforts will focus on establishing a BLB detection model using sensitive spectral features and scaling up using UAV and satellite hyperspectral imagery. We will further explore machine learning algorithms optimized for small datasets and effective feature selection methods to mitigate high-dimensional data-induced overfitting and multicollinearity.