The trend of applying deep learning to images captured by drones continues with a new research project to come out of Kyoto University in Japan, which allows researchers to categorize trees automatically.
Although UAVs have been used for tree classification before, the innovative leap comes from the ability to use affordable off-the-shelf hardware and software, and still maintain a scientific level of accuracy. A DJI Phantom 4 with a 1/2.3 CMOS sensor along with DroneDeploy’s application were utilized for image capture and flight routing. An orthomosaic photo and digital elevation model were made in Agisoft Photoscan Professional and calculations were done with ArcGIS to separate tree crown geometry.
Researchers chose a supervised learning model to train their algorithm, which means humans did the initial category sorting of tree types rather than the computer learning from scratch. As an initial proof of concept, only a few tree types were chosen, but the potential to scale this methodology up to large-scale forest biome tracking is clear and within reach. Traditionally, this type of research would be conducted with LiDar systems, which are cost prohibitive for researchers who do not have a proof of concept or a pressing ecological need to justify their budget or grant requests. The effective use of a simple RGB sensor and a consumer-level drone represents a shift in the availability and complexity of the hardware required to conduct flora classification studies.
This research is part of broader effort by Kyoto University to utilize drones in agricultural research that started this April.