A new system to record growth data for fruit vegetables has been developed. Nodes of crop plants are tagged with a QR code, allowing researchers to record growth events using a tablet PC. Growth data are stored in a database, and tagged with the date and time. As an implementation case, differences in growth of cucumber caused by carbon dioxide gas and mist application were recorded by using the new system. The system allowed acquisition of in-depth data (such as the location where a fruit was harvested, the number of days from bloom to harvesting, and the number of flowers that did not produce fruit) which were previously difficult to acquire, as well as the yield per plant, facilitating analysis. The new system also made it possible to illustrate the growth of individual crop plants, from planting to harvest, as an animation.

Field surveys of rice planthoppers that are important pests of rice in Japan (Nilaparvata lugens, Sogatella furcifera, and Laodelphax striatellus) are conducted by expert researchers in prefectural pest-control stations. Surveys for classification and counting by species and developmental stage involve visual inspection of individual planthoppers captured on sticky boards. This work requires special skill and takes considerable time, especially during severe outbreaks of rice planthoppers, and an alternative method for rapidly counting rice planthoppers is required. We aimed to develop an automatic system for classification and counting of rice planthoppers using an image-based object-detection technique that uses deep learning. Sticky boards were used to capture planthoppers in paddy fields and in the laboratory. High-resolution images of the boards were acquired using a flatbed scanner. Eighteen classes of annotations were applied to these images, and data sets for deep learning were prepared. We used You Only Look Once (YOLO) as the deep learning algorithm for object detection and verified the classification accuracy for each data set and each class combined with learning conditions, such as the image input size. Following the training process, the algorithm was applied to a test dataset, and the highest mean average precision was 0.79 and the corresponding F1-score was 0.88.