Fruit cultivation requires many skills, which are difficult for new farmers to learn. Although the harvesting and sorting of fruit are essential tasks for new yuzu (Citrus junos) farmers, learning these tasks presents challenges, notably with respect to wounds caused by thorns during harvesting and the lack of teaching time for sorting fruit during the busy harvest season. Here, we visually assessed learning support points for these tasks and verified the effects of learning content at these points. First, we extracted 10 points related to harvesting, including “actions”, “decisions”, and “results”, and 6 points related to the sorting of fruit from videos, eye tracking data, and interviews with expert farmers, and used them to assess the work of intermediate or beginner farmers. As learning content, we prepared 21 quiz-type questions and a video for harvesting, and 528 questions and a video for sorting, which focused on decision points where there were differences between experts and intermediate or beginners. As a result of learning, inexperienced workers had a significant reduction in wounds caused by fruit shears and thorns during harvesting and a significant increase in the proportion of fruits sorted in a single inspection. Our findings reveal that using the learning content improved the learners’ knowledge of decisions in the harvesting and sorting of fruit.

In recent years, a shortage of farm workers in green onion production has become a serious problem. Therefore, there is a need to further mechanize preparation. Current machinery cannot remove all unwanted leaves at once, requiring subsequent manual removal. To reduce this secondary processing, it is effective to align the nozzle with the uppermost branch position. It is necessary to recognize the branch position of each green onion and feed the plant into the machine with the branch position aligned. Detecting the branch position by image recognition and automatic alignment of the nozzle for preparation could improve the accuracy of preparation. In this paper, we propose a method for detecting the branch position by extracting a particular oblique line at the branch. The method is designed for implementation on low-power edge devices and uses a by lightweight edge detection algorithm based on image processing. On a Raspberry Pi 3, it achieved a correct detection rate of 90.6% and a processing time of 455 ms. This result shows that our method is effective for detecting the branch position and may be applied to real-world use.

Since 2018, disease of sweet potato foot rot (diaporthe destruens) has been rampant in southern Kyushu, causing serious problems due to the decrease in yield caused by the withering of sweet potato plants and the rotting of tubers. As its appearance in Japan is recent, preventive countermeasures used overseas such as the selection of resistant cultivarsand pesticide spraying have been used and have shown some effectiveness. However, in areas where the damage is severe, it has proved difficult to suppress with these countermeasure alone. We conducted a field survey in areas of southern Kyushu where the disease has caused severe damage to investigate its incidence in 36 fields and the use of countermeasures by farmers, and used the collected data and network analysis to examine the effectiveness of combinations of countermeasures. We found that damage was suppressed in fields where all three countermeasures were implemented. Firstly, resistant varieties were introduced, secondary residual treatment of decayed potatoes after harvest and crop rotation were implemented to reduce soil residual risk, and thirdly drainage measures were implemented to reduce flooding due to rainfall. In regions where sweet potato foot rot disease is severe, measures combining these countermeasures should be prioritized.

As robotic systems become increasingly integrated into plant phenotyping processes, the quality of images captured plays an increasingly crucial role in accurate data collection and analysis. Here we address the challenge of assessing and enhancing image quality in the context of plant phenotyping robot using a pan-tilt-zoom camera. We present a data-driven approach using machine learning, specifically the Random Forest classifier, to classify both blurred and sharp images. Our method involves feature extraction, data preprocessing, hyperparameter tuning, and cross-validation. The resulting model demonstrates promising performance as indicated by its accuracy, precision, recall, area under the curve (AUC), and feature importance analysis. Notably, our results support a highly accurate classifier, achieving a correct classification rate of 95% for sharp images and 92% for blurred ones, a receiver operating characteristic curve with an AUC of 0.93, and a precision-recall curve with an average precision of 0.91. Shapley Additive Explanations analysis identifies “edge density” and “mean gradient magnitude” as influential to the classifier, offering valuable insights for future feature engineering and model refinement. The classifier has a short inference time (2.8 s) on a Raspberry Pi 4B computer, both improving the quality of captured images and automatically eliminating blurred images. By enhancing image quality assessment, this research improves data reliability and the overall effectiveness of plant phenotyping robots. We discuss the implications of these findings and their practical relevance and suggest directions for future research.

Image processing techniques based on neural networks enable highly accurate discrimination in fruit and vegetable sorting. However, conventional classification models may not be sufficiently accurate for grading based on fine features, such as in the simultaneous output of high estimated probabilities for multiple distant grades. Here we propose a classification model combining conventional classification and grade regression for grade discrimination and verified its effectiveness by using the grade discrimination of onion peelings as a test case. The model reduced misclassification to distant grades without decreasing discrimination accuracy, relative to conventional classification and regression models.