DETECTING GANODERMA BASAL STEM ROT DISEASE ON OIL PALM USING ARTIFICIAL NEURAL NETWORK METHOD

. The oil palm tree is one of the essential plants with a major contribution to the Indonesian economy but is also vulnerable to pathogen infection, such as Ganoderma . Ganoderma boninense is a group of polyporous fungi which is responsible for Basal Stem Rot disease. The disease is extremely serious and easily spreads, posing a signiﬁcant threat to the economy, so early detection of the disease becomes vital. However, the current detection techniques for the disease are expensive and time-consuming; hence, they are not ideal for large plantation areas. The development of image processing technology could be utilized to predict Ganoderma infection, using the images that are captured by a drone. This research aims to predict the spread of Ganoderma infection, in the oil palm tree plantation area in North Sumatra, Indonesia, by utilizing image processing and Artiﬁcial Neural Network methods. Our model results showed the prediction accuracy (with Green color) was 73,8%. In addition, we also showed the distribution of Ganoderma infection in the area: score 0 was 229 trees, score 1 was 295 trees, score 2 was 112 trees, score 3 was 238 trees, and score 4 was 23 trees. Overall, our research provided a non-destructive method to detect Basal Stem Rot disease in the oil palm plantation sites.


INTRODUCTION
The oil palm (Elaeis guineensis Jacq.) plantation sector in Indonesia has grown very rapidly in the last 10 years. In 1998, the area of oil palm plantations was 2.7 million ha, with a production volume of 5.6 million tons of CPO (crude palm oil), and the export volume of palm oil was 852,843 tons with a value of US $ 333,866,000 [1](BPS-Statistics Indonesia, 2020). Meanwhile, in 2003, the area of oil palm plantations reached 5.06 million ha, with a CPO production volume of 9.6 million tons or an average of 1.8 tons per ha/year, and the export volume of palm oil was 6.333 million tons with the value of US$ 2.092 billion [1]. The composition of ownership of the oil palm plantation area consists of smallholder plantations at 29.7%, PT Perkebunan Nusantara (Indonesian state-owned enterprise) at 13.2%, and large private plantations at 57.1% [2].
In 2008, the area of oil palm plantations in Indonesia was almost 7 million ha. Like other crops, oil palm is also susceptible to various diseases [3,4]. One of the most common diseases that affect oil palm is basal stem rot disease (BSR), which is caused by Ganoderma [5,6,7,8].
It is the most serious disease in oil palm, especially in Malaysia and Indonesia, the two largest palm oil-producing countries in the world [5,7,9].
BSR was first described in 1915 in the Republic of the Congo, West Africa [10]. In Malaysia, this disease used to infect old oil palms that were more than 25 years old, which would indeed be replanted so that BSR was considered not economically important [11]. However, by the 1960s, when oil palm began to be considered a leading plantation crop, BSR disease continued to increase when even younger oil palms (10-15 years old) were infected [12,13]. Recently, Ganoderma attacked oil palms aged 12-24 years and even plants aged 4-5 years, especially in replanting areas that were originally planted with coconut plants [11].
The attack of BSR in Indonesia was initially low in oil palm plants aged 7 years, then it increased by 40% when oil palm plants reached 12 years of age [14]. In the fourth generation of crops, the attack of BSR occurred earlier, namely on the older plants 1 to 2 years [14]. BSR could also attack oil palm seedlings, presumably because the pathogens that cause the disease are increasingly spreading on land that is often rejuvenated [15]. DETECTING GANODERMA BASAL STEM ROT DISEASE ON OIL PALM 3 BSR disease causes low palm oil production and a decrease in fresh fruit weight. The damage caused by the disease can reach 80% to 100%, and it can even cause death in the affected plants [16,17]. BSR is a threat to various oil palm plantations in Indonesia, especially in plantations that have undergone repeated rejuvenation [15]. A previous study reported that the more often oil palm plantations are replanted or in oil palm plantation areas previously planted with coffee, rubber, or other plantation crops, the lower the diversity, abundance, and distribution of biocontrol agents found [18]. The decrease in the presence, diversity, and abundance of antagonist agents could increase the incidence of BSR.
Given the effect of BSR disease on oil palm trees, detecting the infected trees is crucial. Until now, several methods of BSR detection have been developed, which include the use of Ganoderma selective medium [19], screening with ELISA (Enzyme-linked immunosorbent assay) and PCR (Polymerase Chain Reaction) [20], integrating of headspace solid-phase microextraction (HS-SPME) technique with gas chromatography-mass spectrometry (GC-MS) [21], and the use of the electronic device that analyzed odor [22]. However, these techniques possess several drawbacks, such as being expensive and time-consuming, which are not ideal for implementation on the large scale [23]. Therefore, developing an efficient and more economical method will be beneficial for detecting BSR infection in large plantation areas.
The development of a Geographic Information System (GIS) together with Remote Sensing (RS) has enabled an effective and efficient way to obtain and analyze geographic information on large surfaces, including oil palm plantation areas [24,25]. Further, machine learning approach, such as Artificial Neural Network (ANN) has been used to classify the ripeness of oil palm fresh fruit [26,27,28], weed identification [29], plant counting [30,31], and oil palm leaves deficiency detection [32]. One of the ANN algorithms is Self-Organizing Maps (SOM), which have the capabilities of clustering, visualization, and classification to produce low dimensional representation data called feature map [33]. The SOM has been used to analyze soil physical properties and the distribution of toxic elements in polluted mining soils [34,35]. Therefore, the SOM has great potential to be implemented for the detection of BSR disease on oil palm trees. 4 HERMANTORO, KURNIAWAN, TRINUGROHO, SUPARYANTO, ISNAN, SUDIGYO, AKBAR, PARDAMEAN Based on these reasons, we captured images of oil palm trees with drones and further analyzed them using the Artificial Neural Network method. The purpose of this study is to predict BSR disease in healthy oil palm trees and map the distribution of oil palm trees affected by the disease.  There are three different data types used in this study, including: 1) Ganoderma disease data by location. This data was obtained based on the results of a survey conducted by the company several months prior to the study, which acted a reference in determining the oil palm infected with Ganoderma.

DATA
2) The image of oil palm plantations captured using drones. Drone DJI Phantom 3 Standard was used for taking photos of oil palms in the air. The Pix4D application was used to set the drone flight routes. Then, Garmin GPS 64s was used to create coordinates for oil palm trees that were affected. 3) The image of plant condition. The Camera Canon 600D was used to obtain images of oil palm trees on the land.  in the study by Nasution et al. [36]. The five-class scoring of Ganoderma attack severity is categorized in Table 1. The results of the manual classification were then analyzed using the Self Organizing Maps (SOM) method, one type of ANN, using the Rstudio application. The goal was to obtain a match between the manual classification results and the SOM analysis results.
R2 was used to evaluate the analysis outcomes. Better results were achieved by increasing R2 value. Finally, the Ganoderma distribution map was generated using the highest scoring feature.

RESULTS & DISCUSSION
3.1. Plotting Sample Points. We first plotted a map with the supervised classification ( Figure   2). In the supervised classification, the training area is selected based on the knowledge of the user. The user can define a limit to find out how close the result with the desired result. This limit is often determined based on the spectral characteristics of the existing training region.
The designer can also design the output. with what was expected because there was no color difference that could distinguish between oil palm trees that had a Ganoderma score of 0-4.

Classification of Sample
Points. Next, we analyzed 42 sample points using the Arcgis 10.3 application to obtain the RGB value of each sample point. The Buffer command was used to form a circle as wide as the oil palm tree canopy from the width of the palm leaf, and the intersect command calculated the RGB value of the existing palm sample points. Then, the scoring of the oil palm tree samples was determined ( Table 2). The result showed that half of the sample points had a score of 0 and 1, which were 10 and 11 sample points respectively.
Meanwhile, the other half of the sample points scored between 2-4.

Model Analysis and Correlation Test.
We then performed a test on the actual/manual classification to find out the correlation between the RGB value to the Ganoderma score. The correlation test of the Red (R) color of actual classification showed that the Red value had a positive correlation with the Ganoderma score, suggesting that the lower the R-value, the lower the Ganoderma score ( Figure 3A). This result contradicted the test of Blue (B) color ( Figure 3B), which indicated a negative correlation between Blue value and Ganoderma score.
Meanwhile, the correlation test of the Green (G) color indicated that the Green value in each sample was random ( Figure 3C). This may be because the Green value in the palm canopy was extremely dominant compared to the other color values. nut tree leaves showed lower results in detecting red leaf blotch (Prunus amygdalus) [38]. The accuracy result for the forward stepwise discriminant classification method that integrates the vegetation indices feature is 59.6%. As a result, our research accuracy is slightly more accurate than the two previous studies. Our approach, which focuses on supervised labeling for a more precise categorization, is comparatively greater than using more sophisticated models and gathering satellite photos.
Then, we also performed a correlation test on prediction classification data. The correlation test of the Red (R) color of prediction showed that the Red value had a positive correlation with the Ganoderma score, suggesting that the lower the R-value, the lower the Ganoderma score ( Figure 3D). This is in contrast to the result of Blue (B) color ( Figure 3E), which indicated a negative correlation between Blue value and Ganoderma score. In addition, the correlation test of the Green (G) color suggested that the Green value did not correlate with the Ganoderma score ( Figure 3F). Overall, the prediction results are in agreement with the actual/manual classification. Red and Blue as parameters have a strong correlation in differentiating immature and mature leaves compared to Green in the study of Barman and Choudhury [39]. The RGB parameter used as the primary color was developed into 17 indexes in this study. The combination of the three in the NRI index (R/R+G+B) shows the highest correlation with tender leaves. A similar association result in Barman and Choudhury's study [39] between tender leaves and the The results identified the sample points with the scores 0, 1, 2, 3, and 4 were 8, 5, 13, 12, and 4 points, respectively. Thus, the accuracy of the prediction was 59,5%, which was lower than the prediction using the Green color. The result suggested that the Green color is essential for predicting the presence of BSR disease. This may be because one of the main symptoms of BSR disease is the loss of chlorophyll in plants, leading to the change of color in plant leaves [23]. Overall, our result agreed with the previous research by Hushiarian et al [40].
In addition to the chlorophyll index, the vegetation index affects the accuracy of the model performance in detecting and classifying diseases and the physical condition of plants. In several studies, feature important analysis shows that RGB parameters individually show lower predictive scores compared to complex color combinations as features. The NIR and Red value as features measurement in the study by Raza et al. [41] in detecting soybean sudden death syndrome. Research by Ciocîrlan et al. [42], the combination of RGB parameters as a complex index shows a higher scoring than the simple RGB index in the Random Forrest model.
In the future, multiple indexes for color normalization methods will be used for measuring the variables. In addition to using sophisticated machine learning models, it is envisaged that using these variables as features can enhance the performance of ANN models. 3.5. Analysis of All Points Using ANN. After confirming that the Green color value is needed to obtain high accuracy, we then analyzed all sample points using ANN (Figure 4).
The result showed that most of the sample points/oil palm trees had a Ganoderma score of 1 (295 trees), while the least sample points had the score 4 (23 trees). In addition, only a quarter of the sample points were not affected by Ganoderma, suggesting that the BSR disease dominated the oil palm plantation area. We further plotted the result from ANN analysis to the map to identify the spread of the Ganoderma attack (BSR disease). From Figure 5, it could be seen that the Ganoderma score was evenly distributed, indicating that the spread of BSR disease was not clumped in certain sample points. It is known that BSR could spread rapidly by physical contact with the infected root [11,43]. Furthermore, the BSR disease could also spread from the basidiospores, which can easily travel through the air [44,45]. Overall, the results showed that the majority of the sample points had been affected by the BSR disease.

CONCLUSION
This research has provided evidence that the machine learning approach, especially the Artificial Neural Network (ANN) method, could be used for analyzing the image of oil palm trees.
The results of the analysis using the ANN method showed the distribution of Ganoderma disease in the plotted map. The result of the correlation test using a linear graph displayed that the lower the Red value, the lower the Ganoderma score. On the other hand, the correlation test of the Blue color displayed a negative correlation between the Blue value and the Ganoderma score. The Green value did not show a strong correlation with the Ganoderma score, which may be due to the dominance of the Green color in the canopy of the oil palm tree. In addition, the ANN analysis had identified the spread of BSR disease in all sample points, which was indicated by the Ganoderma scores: Ganoderma score 0 was 229 trees, score 1 was 295 trees, score 2 was 112 trees, score 3 was 238 trees, and score 4 was 23 trees. Overall, our research provided efficient and non-destructive ways to detect the spread of BSR disease in oil palm plantation sites. Further study should be undertaken to improve the accuracy of the prediction so that it can be implemented in a larger area.