Estimation of Longan Yield Using Bayesian Belief Network Modeling in Phrao District, Chiang Mai Province

• Main Authors: วรวีรุกรณ์ วีระจิตต์, Vorraveerukorn Veerachitt
• Other Authors: ชาญชัย แสงชโยสวัสดิ์
• Format: Theses and Dissertations
• Language: English
• Published: เชียงใหม่ : บัณฑิตวิทยาลัย มหาวิทยาลัยเชียงใหม่ 2018
• Subjects: Longan; Bayesian belief
• Online Access: http://cmuir.cmu.ac.th/jspui/handle/6653943832/46032
• Institution: Chiang Mai University

In this study, spatial data maps were created which could be adjusted for accuracy and updated as needed. Especially they were developed from the originally official land use map for improvements into accurate detailing of the land at the orchard level. With these modified official maps for greater contents, the entire district area could be understood with detailed accuracy, and every location of longan orchards could be mapped out, including the age of the orchards and what small water resources were being used for the orchards. All of these data were used to create a Land Mapping Unit (LMU) for longan production by overlaying the data in three layers. These three layers of data were the ages of orchards, the methods used for watering the orchards, and the land slope. The LMU contained 18 map units, which were important areas of interest for this study. The LMU was used in the process of randomly choosing farmers to interview as well as farms to sample for photographic identification of the fruits in the trees’ canopies. When choosing which trees to photograph, 60 trees were sampled from 30 different orchards. Each tree was photographed six times by using 1 x 1 meter square frames of plastic pipe in the tree’s canopy. The photos were taken with a digital camera and were originally in the JPG format. So altogether, there were 360 photos taken. These photos were categorized into five groups of lighting conditions: sunny with less light, sunny with medium light, sunny with bright light, cloudy with less light, and cloudy with medium light. All of the photos were converted into the RGB format, and the statistics from their layered data were analyzed, especially concerning the relationships among the data in each of the five groups. These statistics and relationships became the defining information for creating the threshold for the identification of the longan fruits within the photos. The identification process was done using the ERDAS Imagine 9.0 program. Each type of object in the photos was identified—a total of six types of objects—in order to make the identification of the fruits easier. The initial accuracy rate of identifying longan fruits in the photos was more than 70%. The BBN model (Bayesian Belief Network Model) was used to predict the yield of longan fruits. In order to create a reliable model, first a flowchart had to be made based on the information gathered from interviews with 150 farmers and from the focus group’s discussions. All of this information was used to create nodes in the flowchart and to understand the relationships between these nodes, which are factors impacting the production of longan fruits. The flowchart’s purpose is to display the reasonable connections between each node so that a better understanding of the factors enables the BBN model to be accurate in its predictions. Three models were designed based on three different sources of information concerning longan yield results. The Type A model’s data originated from questionnaires which farmers filled out. The Type B model’s data originated from first-hand gathering of information on-site during harvest time. The Type C model’s data originated from the image analysis of the photos taken on-site before harvest time. These three types of data were used to create a Conditional Probability Table (CPT) of the longan fruit yield. The results show that the Type B model has the lowest value of RMSE at 20.2, while Type A and Type C have a value of RMSE at 44.7 and 73.3 respectively. The average predicted yields from these three models were not significantly different from the actual level. The Type B model had a difference of 13 kg/rai, which is a difference of 1.5%. The Type A and the Type C models differed from the actual results by 123 kg/rai (a difference of 16.5%) and 87 kg/rai (a difference of 11.2%) respectively. The model with the best results was used to create a map of the distribution of longan production throughout the district. It was based on the LMU. The model was also used to calculate the total output of longan fruits for Phrao District which turned out to be 35,316,227 kg in a total longan area of 40,400 rai. Keywords: Bayesian Belief Network (BBN), longan, image analysis, map of yield predicted, longan production.