Evaluation of multiple In Situ and remote sensing system for early detection of Ganoderma boninense infected oil palm

Ganoderma boninense is a causal agent of basal stem rot (BSR) and is responsible for a significant portion of oil palm (Elaeis guineensis) losses, which can reach USD 500 million a year in South East Asia. At the early stage of this disease, infected palms are symptomless, which imposes difficulties in detecting the disease. In spite of the availability of tissue and DNA sampling techniques, there is a particular need for replacing costly field data collection methods for detecting Ganoderma in its early stage. This study evaluated the use of insitu and remote sensors to early detect the Ganoderma infected oil palms before the visual symptoms are manifested (mildly infected). The study was conducted in Machap sub-district belonging to United Malacca Berhad located in Melaka, Malaysia (2.402° N and 102.327° E) (WGS 84 coordinate system). Initially an experiment was carried out to determine the best insitu sensors that could be utilized for early detection of Ganoderma in oil palms. During the field experiments, leaf samples of healthy (T1), mildly (T2), moderately (T3) and severely-infected (T4) palms were measured using a Minolta SPAD-502 chlorophyll meter and a SC-1 leaf Porometer to obtain relative leaf chlorophyll content and stomatal conductance, respectively. Afterwards spectral reflectance readings data were acquired using a GER 1500 spectroradiometer from 1016 spectral signatures of foliar samples in four disease levels (T1 to T4) and 2 fronds (9, 17). Various artificial neural network (ANN) architectures were applied to the datasets to verify the proficiency of various combinations of input variables, learning optimization methods and different numbers of neurons on the hidden layer by MATLAB 2014a software. The neural network chosen in this study was multi-layer and back-propagation (BP) due to the ability to learn and determine nonlinear combinations. 70.0% of data were assigned for the purpose of training the network, while the remaining 30.0% of data were allocated for testing model accuracy.Subsequently in imaging processing study, 287 oil palm samples were classified into three disease levels (T1 to T3) using ANN, whereby the principle of the classification is to seek for the most representative image configurations and network properties while adjusting for the best canopy circle radius, threshold limit, best neuron numbers of hidden layer and the best mean and standard deviation values from different combination of spectral bands (green,red, and NIR bands) from CIR images obtained from a unmanned aerial vehicle (UAV). Simultaneously, the number of hidden neurons and termination error were optimized given various classification input in order to correctly classify the imaged palms to their corresponding severity classes. For this purpose, the dataset was randomly split into three sets, 60.0% for model training, 20.0% for model validating, and 20.0% for model testing. In the second stage and for improvement of image processing study, support vector machine (SVM) classifier was performed on UAV and Pleiades imagery to identify early Ganoderma infected oil palms. In the first phase, spectral features and structural features were extracted for feature extraction. In the spectral features part, the descriptors include red (R), green (G), blue (B), near-infrared (NIR) digital numbers and a vegetation index (VI) was considered. Statistical parameters like average, variance and grey-level co-occurrence matrix (GLCM) was set as a structural feature which provides several statistics information about the texture of an image. In the next phase, the SVM classifier was trained to achieve the best classification using training data and test data integrated with selected features. The results and consequences of this study showed that the chlorophyll meter, leaf porometer and spectral indices from the spectroradiometer, mNDVI, GNDVI and VOG1, were found beneficial to differentiate between T1 and T2. Nonetheless, the combination of VOG1-stomatal conductance obtained from frond 9 and 17 could discriminate the T2 palms from the T1 ones with accuracies ranging from 66.67% to 73.68% regardless of time of measurements. The results obtained from the spectroradiometer analysis presented that the healthy oil palms and those which were infected by Ganoderma at early stage (T2) were classified satisfactorily with an accuracy of 83.33%, and 100.0% in 550-560 nm, respectively, by ANN using first derivative spectroradiometer data. The results further indicated that the sensitive frond number modelled by ANN provided the highest accuracy of 100.0% for frond number 9 compared to frond 17. The results acquired from the UAV classification in the third study indicated that classification error of 12.29% was achieved which generated by the ANN network by 219 neurons, green and NIR bands, canopy circle radius of 35 pixels and 1/8 threshold limit. The total classification accuracy for training and testing the dataset was 97.52% and 72.73%, respectively. Finally, the findings from fourth study showed that the best prediction results of using SVM classifier were obtained from the UAV image with an overall accuracy of 68.28% compared to Pleiades with an overall accuracy of 64.52%; also the early Ganoderma infection (T2) could be detected with an accuracy of 64.07% and 64.49%, respectively. Even though at first glance the classification accuracy was moderate, the level of details provided by the imageries suggests that the accuracies were acceptable. In conclusion, for early detection of Ganoderma, better accuracies were derived from the spectroradiometer which is a destructive and ground based method and still requires individual leaf sampling. This method is still considerably time-consuming and laborious compared to the rapid, nondestructive approach of canopy reflectance using UAV or satellite imagery. The latter, nonetheless, gained reasonable accuracy and is appropriate for field applications involving mass screening of oil palm plantations that are commonly cultivated in thousands of hectares in seeking for potentially infected individual palms. This study concluded that remote sensing approach combined with data mining approaches such as ANN algorithms have great potential in monitoring vast plantation areas in a rapid and inexpensive manner