Canarium Ovatum Recognition utilizing Mask R-CNN and Lightweight Unmanned Aerial Vehicle

Abstract

Canarium Ovatum, commonly known as Pili Tree in the Philippines, was known for its potential for food production and other utilization possibilities. Pili was widely grown in large areas since they grew taller and needed enough spacing for their crown diameter. Currently, monitoring and inventory of these Pili trees are done manually, which is expensive and painstaking to be implemented. On the other hand, the proliferation of Remote Sensing technology is now being applied to areas of agriculture and forestry. However, although reliable and cost-effective, this technology is still expensive for typical utilization. This study explores the performance of the Mask R-CNN model on orthomosaic data obtained from a light Unmanned Aerial Vehicle (UAV). To achieve this, a Mask R-CNN model was trained using the training datasets from the pre-processed orthomosaic obtained using a DJI Spark, its built-in camera, and Pix4D. A system user interface was also developed to facilitate the loading of the data and validate the model's performance. The initial output of the model was then verified using the data resulting from splitting the orthomosaic and comparing it with the ground-the-truth data. Using the developed interface in python, the model can recognize the Pili tree with an accuracy of 89% and an f1-score of 0.9231.