Researchers at Kyoto University have used a drone carrying a camera to take aerial photos of trees and classify them automatically using a deep-learning algorithm.
Often, ecological scientists and land use managers requiring tree classification on a large scale have to rely on expensive, specialised sensors in order to gather and analyse accurate data. However, in a paper published late last month, the researchers described a process using a simple UAV that may reduce costs and increase efficiency of these processes.
In the paper, they explain how they combined a consumer-grade drone with a publicly available deep learning package. Flying the drone over a forest in Kyoto, Japan, the researchers were able to segment the aerial images into individual tree crowns.
While many such methods utilise expensive imaging technology such as multispectral sensors to enhance the data, the researchers used a low fidelity camera that took very basic RGB photos.
Using the RGB images, the researchers were then able to tweak the deep-learning algorithm to recognise the tree crowns and divide into seven categories. Of these categories, six were types of trees and a final category called “others” was included for images that captured bare land or buildings.
The results of the study were highly accurate, with an 89% success rate.
The success of the project is being hailed as significant due to the simplicity of the technology and the potential savings for those involved in forest research and land management. The method could negate the need for researchers and forestry managers to invest in more expensive systems with expensive multispectral or LiDAR payloads.
It is also possible that the relatively simple technology could pave the way for similar systems to be trained for other applications such as disaster response, pipeline inspection or other industries requiring fast, cheap identification of objects from the air.