Researchers use Phantom 4 and AI to identify tree species in Japanese forest

A research team used an off-the-shelf DJI Phantom 4 and AI software to identify 7 different tree species in a forest in Kyoto, Japan. After some tweaking of the algorithm, the drone and software were able to identify the tree in the forest with an 89% accuracy. This is notable because the team was able to achieve these results with a consumer drone and regular RGB images instead of expensive hardware, such as multispectral imagers and thus represents a significant cost-savings for forest researchers and managers.
 
The ability to automatically identify the trees in a forest has been a dream for many scientists and land use managers as it is useful for biodiversity assessment, monitoring of invasive species, wildlife habitat mapping, and sustainable forest management. Typically expensive equipment, such as airborne hyperspectral, multispectral, and LiDAR sensors is required, but now a Japanese research team has been able to achieve satisfactory results (89% accuracy) with an off-the-shelf DJI Phantom 4 drone and AI software.
 
Especially when compared to a manned aircraft, drones are a very cost-effective and easy-to-use tool. UAV’s can fly much closer over the tree canopies and capture higher resolution images than is possible with a manned aircraft. In their research report, it was mentioned that:
 
“The combination of UAVs photography and deep learning is expected to have a high potential for classifying trees even if we use consumer grade digital camera. And also, this machine vision system will be a cost-effective and usable tool for forest managements.”
 
After some tweaking of the algorithm, the team was able to identify 7 different tree species in the Kyoto forest with an accuracy of 89%. However, the team also points out that one of the reasons for the great results is that:
 
“We picked up training and test images from the same area and the same time. Tree shapes are thought to be different in different environments, and leaf colors and illuminations are different at season and weather. Making a good use of tree shapes (or DEM) and seasonality of leaf colors will improve classification accuracy, but generally these properties may have a bad influence for simple machine learning. Considering practicability, versatile model which is trained images of various site and time is desired in the further study.”

---