Early prediction of the best harvest time and making a precise harvest plan is vital in tea cultivation, because harvest time has a substantial impact on tea quality. However, existing models are difficult to predict the best harvest time, even 2 weeks in advance. In addition, budding time needs to be known. Here we developed a linear multiple regression model to predict tea harvest time by using the integrated values of weather data before budding time (1 month before harvest time). The harvest times of the cultivars ‘Yabukita’, ‘Sayamakaori’, and ‘Okumidori’ planted in Uji City, Kyoto Prefecture over a period of 10 to 19 years were used. Climate components—daily mean relative humidity, daily precipitation, daily daylight hours, and one of the daily mean, maximum, and minimum air temperature—were selected via the stepwise method. All weather data combinations within 30 days from 1 March to the average budding time were used as the integration period. The best prediction model for each cultivar was selected, with the smallest mean absolute error (MAE) between the predicted and recorded harvest times of the validation data based on the leave-one-out cross-validation test (MAE_test). As a result, the MAE between the predicted and recorded harvest times ranged from 0.8 to 2.0 days, which was accurate enough for practical use. In addition, the MAE_test ranged from 1.2 to 2.3 days—less than that of our previous model trained with the integrated value of weather data after budding time. The best time to harvest tea could, therefore, be predicted approximately 1 month before harvest.

We examined the conditions required to achieve both scale expansion and yield increase by using farming data and smart agricultural machinery in large-scale rice-paddy management. Actions from 2020 to 2022 by two large-scale paddy farm management corporations in Ibaraki Prefecture were examined from the perspective of data-driven agricultural production, rice yield per 0.1 ha, rice production cost, and operation of autopilot rice transplanter. These two corporations actively introduced smart agricultural machines and hired new employees. They expanded their operating areas by more than 40 ha each in 3 years, but their rice production costs per 0.1 ha and per harvest were reduced. For timely and appropriate cultivation management during the phase of rapid expansion, these two corporations implemented data-driven management improvements by building field-specific datasets and predicting growth by using a cultivation-management support system. They also readjusted their planting order and cultivation methods to suit each cultivar. As a result, even though the area of cultivated fields under their management increased rapidly, they were able to manage cultivation in an appropriate and timely manner. To increase yields in the phase of scale expansion, it is necessary simultaneously to improve land conditions by consolidating neighboring fields, improve work efficiency by operating smart agricultural machinery, and improve cultivation management by analyzing collected farming data.

From a field-management perspective, monitoring flower buds and fruits in blueberry cultivation is essential. However, blueberry cultivation in Japan is mainly small-scale and manual, which hinders real-time monitoring of the entire field. Here, with the ultimate aim of automating monitoring of blueberry fields, we examined the feasibility of estimating the number of flower buds and fruits per bush by using object detection through deep learning, specifically employing YOLOv5 and YOLOv8 models. First, we detected and classified normal and frost-damaged flower buds and unripe and ripe fruits at various resolutions using a DSLR camera. Then, we estimated the numbers of flower buds and fruits in the images by using a simple regression formula. The results showed high accuracy of the automatic counts of normal buds, ripe fruits, and unripe fruits in the images but low accuracy for frost-damaged buds. For flower-bud detection, the highest average precision (AP) index was 0.755 and the lowest regression error index mean absolute percentage error (MAPE) was 10.79%. For fruit detection, the highest AP was 0.815 and the lowest MAPE was 23.84%. Furthermore, we tested the possibility of estimating the total number of fruits, including those hidden behind leaves, from the number of fruits visible in the images. We found a positive correlation between the number of fruits in the images and the total number of fruits (r=0.96, p<0.001), with a simple regression error MAPE of 21.2%. Also, the degree of leaf coverage markedly affected the estimation error. However, we found that the imaging direction was inconsequential and image collection from both sides of the row was unnecessary. Finally, we examined potential field applications of this automatic flower bud and fruit counting and classification technology, such as early yield prediction (during the flower bud phase), understanding the spatial variation of pollination effects by bees, and determining the optimum harvest time and harvest sequence.

We developed a decision-making tool for broccoli clubroot disease control. The tool uses the LAMP method for genetic diagnosis to visualize the density of broccoli clubroot disease in each field and can provide prescriptions for control measures. The tool allows farmers to check items such as clubroot disease bacterial density, diagnostic result type (bacterial density), soil pH, and remedial measures for each field on the “Aisaku” farm management support platform. The visualization tool has enabled farmers in the Shimabara Unzen study area and Japan Agricultural Co-operative’s agricultural advisors to share information and implement countermeasures. After the introduction of the tool, 60.9% of producers became aware of the need to implement measures based on information from the tool and farm management guidance. The results showed that over the 2-year period, pesticide purchases decreased by 33.6%, lime material purchases increased by 41.2%, and decoy-plant seed purchases increased by 591.3%, suggesting the effectiveness of the tool’s implementation. A visualization tool that enables field-specific implementation of countermeasures on the basis of fungal density can be a powerful decision-making tool for countermeasures that can reduce the excessive use of pesticides and facilitate implementation of sustainable agriculture.

Dystocia among dairy cows results in economic losses. Such losses can be prevented, and the survival of newborn calves can be ensured, by predicting the degree of calving difficulty in cows. Although pelvic cavity measurements have been commonly used to predict calving difficulty, these are difficult to obtain because measurement requires considerable time and effort. Hence, a simpler and more practical alternative to pelvic cavity measurements must be developed. Here, we applied the deep learning algorithm of the convolutional pose machine (CPM), a model for estimating joint positions, to predict the degree of calving difficulty in Holstein dairy cows by estimating key points on the lower limbs and pelvis from images. Side and back view images were obtained. The image data were augmented for training and validation to determine the skeletal key points. The data were augmented to Training (n=189,125) and Validation (n=48,790) sets to learn the skeletal key points. A total of 23 key points related to the pelvic shape of the cattle were identified, and the percentage of correct key points was 0.89. Therefore, skeletal key points in cattle can be estimated by using CPM, and the pelvic skeleton recognition system developed here can be used to construct a system for estimating the degree of calving difficulty on the basis of images.