DEVELOPMENT OF TREE DETECTOR MODEL AND WEB CLIENT FOR ARTIFICIAL INTELLIGENCE SYSTEM TO DETECT TREES WITH THE POTENTIAL TO DAMAGE POWER LINE

• Main Author: Ardelia Hanifah, Karisa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55557
• Institution: Institut Teknologi Bandung

According to data from Unison Group New Zealand, 20% of unwanted power outages occur due to vegetation growing too close to power lines. In Indonesia itself, there has been a case of blackout caused by the Sengon tree growing through the Right of Way (ROW) of electricity network. One of the causes of this incident was the difficulty in managing the vegetation around the power line, especially in remote areas. An Artificial Intelligence (AI) system to detect trees that have the potential to harm the power line was developed as a solution to this problem. By using object detection technology, AI will detect trees and power lines in the image from the UAV and satellite imagery. Then, AI will calculate the distance between the trees and the power line, then assign a box with a certain color to each tree according to the calculated distance. This AI is integrated with a web client, so that users can easily access this AI. The focus of this final project is the development of a tree detection model and a web client. Development is carried out by conducting literature studies, design making, implementation, and testing. Tree detection will be performed using the DeepForest library. Transfer learning from the DeepForest prebuilt model will be performed on the associated data to optimize tree detection. The performance of tree detection model will be measured by mean Average Precision (mAP) parameter. First, the prebuilt model will be tested with varying patch size parameter to determine a good patch size value to use. Then, transfer learning is carried out and the performance of each resulting model will be compared. The model with the best performance will be selected. Web client testing will be carried out by functionality test and Technology Acceptance Model (TAM) survey. In the survey, users will judge how easy to use and informative the web client is. From this research, patch size values that can be used to produce a good tree detection, a better model for detecting trees from satellite imagery, and a web client with proven functionality and convenience are obtained.