Harvesting accounts for the majority of working hours in cabbage production, and labor-saving during harvesting is an important issue. A cabbage harvester is required to solve this issue, and it is therefore necessary to be able to perceive and assess the uniformity of cabbage growth in fields in order to harvest most efficiently. Remote sensing using drones has been used to visualize plant growth in many crops, especially rice; however, little research has been conducted on cabbage using this technology. We investigated the use of aerial photography using drones and image analysis to evaluate cabbage growth at the individual level.

Using image analysis with structure from motion and multi-view stereo (SfM-MVS) and geographic information system (GIS) software, we calculated the horizontal projected cabbage area and vegetation indices, including the normalized difference vegetation index (NDVI) and normalized green-red difference index (NGRDI), for each individual cabbage. The extraction rate of cabbages was the highest at flight altitudes of 20-30 m with a ZenmuseX3 camera, generating ground pixels 0.9-1.3 cm in size. The optimum flight altitude for operating efficiency was found to be 30 m. Correlation analysis of the harvested head weight suggested that the horizontal projected area and Day 15 after plantation were a suitable growth diagnosis index and timing, respectively, for estimating the harvested head weight. In addition, correlation analysis of the horizontal projected areas and observed leaf areas showed a significant positive correlation, with an R value of 0.99. These results suggest that the observed leaf areas can be estimated using drone imagery. We also created field maps to visualize the growth of cabbages at the individual level, and to precisely identify the positions of cabbages with delayed growth. These results indicate that remote sensing using drone imagery can be useful for evaluating cabbage growth in fields.

The National Institute of Information and Communications Technology (NICT) has been observing the ground surface with the Pi-SAR2 airborne synthetic aperture radar, and has been studying methods to utilize the observation data effectively. Deep learning, a type of machine learning, is a method that shows high performance in the field of image classification and recognition, and that has also been actively studied in the field of remote sensing. In this paper, we report the results and verify the accuracy of a deep learning approach to land cover classification for high-resolution and full-polarimetric data observed by the Pi-SAR2.