Individual Tree Crown Detection Using Deep Learning Model in Tropical Forest

Zhafri Roslan, Zalizah Awang Long and Roslan Ismail

The task of identifying individual tree crown in tropical forest is still a challenge. The impact of identifying each tree could benefit in sustaining the forest for future generation. Remote sensing techniques and LiDAR data are known to be the primary contributor to the field of tree crown detection. However, with the emergence of advance technologies such as highresolution satellite cameras, Unmanned Aerial Vehicle (UAV) drone, and advance deep learning networks, the detection of individual tree crown in tropical forest could be achieved. This paper proposes a supervised deep convolutional neural network pipeline using a powerful object detection model called RetinaNet. An unsupervised K-Mean clustering algorithm is applied to optimize the anchor ratios and the result showed an increase in precision. Furthermore, a logcosh regularization will also be utilized to compare the performance of the prediction compared to a smooth-L1 loss. The performance of several well-known classification backbones is used to analyze the result. The model produced a higher F1 score of 0.6458 with more than more than 50% of the trees are correctly predicted. The result from this study could be used to further improve future tree detection model.

Volume 12 | 08-Special Issue

Pages: 239-255

DOI: 10.5373/JARDCS/V12SP8/20202522