A continual monitoring is necessary for quantitative evaluation of ecosystem service of the seagrass beds. In general, seagrass monitoring had been conducted by remote sensing using satellite or aerial images and by field survey. Recently, use of drone in field science is increasing. Drone can take much higher resolution images than satellite images. For mapping seagrass in such fine resolution, traditional classification methods such as pixel- or object-based classification may not be suitable because they only use limited information such as bands value. In contrast, deep learning image analysis can detect many features from input images, and thus is expected to be useful for mapping seagrass from drone images.
In this study, we compared the accuracy of the three methods (pixel-based classification, object-based classification and application of deep learning). Field surveys and drone photography were conducted in Lake Saroma, Hokkaido.
Comparison of different classification methods revealed substantial increase in accuracy when deep learning was applied. Deep learning also classified two seagrass species (Zostera marina and Z. japonica) correctly.
Our results showed that pixel- and object-based classification will no longer be the best option and application of deep learning is suggested for analyzing high resolution images by drones.