Deep Learning for Tree Crown Detection In Tropical Forest

摘要

The task of detecting trees in tropical forest has been a challenging task for decades as it is still not ready for operational purpose. Remote sensing and LiDAR technology has been the dominant research field in adhering the problems of tree crown detection. However, the advancement of deep learning technology could bring new results. This study proposed a deep learning approach using object detection algorithm to detect tropical trees based on tree crowns. The method used state-of-the-art object detection model and K-Means clustering for management of anchor ratio to improve the prediction of the model. Hand-annotations labelled used on the aerial imagery for supplementing the training data. The model produced more than 70% mean Average Precision (mAP) using a small training dataset. The result could open doors to further improve the model for detecting the trees especially in complex canopy forest like tropical forest.