Early detection of Ganoderma boninense in oil palm seedlings using hyperspectral images and machine learning classification techniques

Basal stem rot (BSR) caused by Ganoderma boninense (G. boninense) fungus is one of the most destructive diseases of oil palm plantations in Southeast Asia that resulted in losses up to USD500 million annually. Besides mature trees, seedlings are also susceptible to G. boninense infection after being transplanted into a plantation. Therefore, early detection, timely prevention and control are crucial because trees with less than 20% infection can still be treated. This study focuses on early detection of G. boninense in oil palm seedlings based on the physical growth of seedlings, spectral reflectance of leaves and machine learning classification. Twenty-eight oil palm seedlings aged five months old were used whereby 15 of them were inoculated with the G. boninense pathogen. The physical growth of the seedlings was monitored for every two weeks, and parameters recorded were fronds count, chlorophyll content, height, and girth. The physical growth did not provide any significant differences between the uninoculated (H) and inoculated (U) seedlings throughout the study, which indicate no BSR symptoms had appeared; however, the H obtained marginally higher measurements in most weeks. After 20 weeks of inoculation, spectral reflectance oil palm leaflets taken from fronds 1 (F1) and 2 (F2) were obtained using Cubert FireflEYE S185 hyperspectral camera with wavelength ranging from 450 to 950 nm. The differences between H and U were observed in the NIR and red-edge spectrum for reflectance and first derivative spectra, respectively. Thirty-five bands were found significant for reflectance and 14 bands for first derivative spectra. The bands were later used as input parameters to develop F1, F2, a combination of F1 and F2 (F12), F1 derivative (F1dev), F2 derivative (F2dev), and F12 derivative (F12dev) classification models, i.e., decision trees, discriminant analysis, logistic regression, naïve Bayes, support vector machine (SVM), k-nearest neighbor, and ensemble. These bands were optimised according to the classification accuracy achieved by the models. The result showed that the acceptable number of bands to develop classification models was 11 bands which obtained accuracies of 100% (F1), 92% (F2), 95% (F12), 97% (F1dev), 90% (F2dev) and 93% (F12dev) which considered the highest of its classes. Overall, 11 bands of F12 provided near good linear SVM model with 95% accuracy and a kappa value of 0.9; it was considered the best model since it did not require complex pre-processing to separate F1 and F2. This information is useful in aerial-view applications when applying an unmanned aerial vehicle (UAV) for image acquisition since both fronds can be clearly seen from the top-view image hence could expedite the detection of the BSR disease.

Text FK 2021 49 - IR.1.pdf Download (1MB)

Actions (login required)

View Item