Utilizing Convolutional Neural Networks (CNN) for Persimmon Irrigation Decision-Making: A Case Study in the Gojo Yoshino Region, Nara Prefecture, Japan

Utilizing Convolutional Neural Networks (CNN) for Persimmon Irrigation Decision-Making: A Case Study in the Gojo Yoshino Region, Nara Prefecture, Japan | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Utilizing Convolutional Neural Networks (CNN) for Persimmon Irrigation Decision-Making: A Case Study in the Gojo Yoshino Region, Nara Prefecture, Japan Atsushi Okayama, Atsushi Yamamoto, Masaomi Kimura, Yutaka Matsuno This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4404121/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 14 Jan, 2025 Read the published version in Environmental Monitoring and Assessment → Version 1 posted 4 You are reading this latest preprint version Abstract This study aimed to develop and evaluate a model for persimmon irrigation decision-making using Convolutional Neural Networks (CNNs) based on leaf image data. Data comprising leaf images collected in the field and corresponding soil moisture measurements were gathered from the Gojo Yoshino region, recognized as the primary persimmon-producing area in Nara Prefecture, Japan's second-largest persimmon-producing prefecture. The findings demonstrate that the constructed CNN model can successfully identify water stress levels in persimmon trees from leaf image data. However, there are limitations to the model's performance and scopes for improving accuracy. The model's capability enables remote irrigation decision-making by integrating field-acquired leaf images into edge devices for on-site processing. When integrated with ongoing developments in remote irrigation systems, this technology has the potential to automate irrigation practices, thereby offering substantial labor-saving benefits. Persimmon irrigation Convolutional Neural Networks Decision-making Remote sensing Automated irrigation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1. Introduction In Japan, persimmons have held a prominent status among the nation's fruits for centuries, with a cultivation history dating back to ancient times. The commercialization of persimmon cultivation can be traced to the 1920s in the Gojo Yoshino region of Nara Prefecture, Japan. Situated in the mountainous core of Japan (Fig. 1 -a), this area has an average annual temperature of 15.3°C and annual precipitation of 1,492 mm, as recorded in 2023 by the Japan Meteorological Agency (JMA). The 1970s saw a significant increase in land development initiatives aimed at facilitating persimmon cultivation across the rugged terrain, driven by the inception of a national agricultural development project in 1974. Consequently, the Gojo Yoshino region emerged as Nara Prefecture's primary persimmon-producing hub, eventually becoming Japan's second-largest persimmon-producing prefecture. Cultivation methods in this region encompass both greenhouse and open-field practices (Fig. 1 -b). To ensure consistent fruit yields, irrigation water is sourced from the Ichinoki Dam (Fig. 1 -c), a national government undertaking completed in 1997. The dam supports an extensive irrigation network, covering a total area of 1,817 hectares, with approximately 1,600 hectares dedicated to persimmon cultivation (Gojo-Yoshino Land Improvement District, 2002 ). In persimmon cultivation, numerous studies have underscored the substantial influence of irrigation quantity on fruit size and yield. For instance, Aoki et al. (1971) conducted a study comparing different treatment zones involving irrigation and nitrogen application for 'Fuyu' persimmon fruit. They observed that the irrigation treatment zone yielded more fruits, with a higher proportion falling into the L-grade or higher category compared to the non-irrigated treatment zone. Similarly, investigations on the multi-cultivation irrigation system for 'Taiaki' persimmon fruit, carried out by Matsuda et al. ( 2011 ), revealed that the irrigation treatment zone produced larger fruits than the non-treated zone. These findings strongly indicate that increased irrigation can significantly augment harvest yields. In the Gojo Yoshino region, irrigation of open fields is typically scheduled during the summer months, spanning from June to September, coinciding with the growth period of persimmon fruits. Each field is assigned specific irrigation times, typically occurring for one hour every three days. To accommodate farmers' schedules and mitigate the impact of peak working hours, irrigation is often timed for early mornings or late evenings extending to midnight. However, due to the scattered distribution of fields across steep mountain slopes, accessing these sites demands significant effort and time from farmers. In an effort to streamline this labor-intensive process, plans are underway to implement a low-power wide-area network (LPWAN) featuring an automatic and remote irrigation system (Fig. 2 ). Distinguishing itself from conventional LPWANs such as LoRa, this relatively new system offers an impressive communication range of 100 km on flat terrain, representing a tenfold improvement. Initial tests of a prototype system have already demonstrated successful communication across a mountaintop spanning approximately 10 km. Communication Standard Currently, persimmon irrigation decision-making relies on the experience and judgment of individual growers who observe the trees in the field. Even with the introduction of the new LPWAN system, which enables remote control and automation of irrigation, the workload of farmers may not decrease if they still need to physically visit the fields to make irrigation decisions. Therefore, there is a pressing need for technology that facilitates remote irrigation decision-making. Soil moisture monitoring stands as a well-established technique for irrigation scheduling. However, the variability of soil moisture across different soil types and locations presents a challenge to its practical application. To address this challenge, the assessment of soil water potential, quantified by the pF value, has been proposed. This metric enables the determination of water holding capacity across diverse soil types and facilitates tailored irrigation scheduling to meet crop-specific requirements. The pF value is defined as the logarithm of the negative pressure (measured in cm of water) necessary to extract water from the soil, as depicted in Eq. (1): pF = log10(− h) (1) Where h represents the pressure head in cm of water. Utilizing the pF value as a benchmark for persimmon irrigation, Kamoda ( 1987 ) proposed optimal pF values tailored to each growth stage. Furthermore, studies, such as that by Iwata et al. ( 2019 ), indicate the effectiveness of soil moisture management below pF 3.0, a threshold considered growth inhibitory. While remote irrigation decision-making based on these values is feasible, maintenance poses challenges due to the necessity of burying sensors, rendering visual inspection of equipment impossible. Moreover, this method incurs costs associated with sensor re-burial in case of malfunctions or calibration. Consequently, attention has shifted towards irrigation decision-making grounded in canopy inspection from above ground. Unlike buried sensors, this approach targets the visible canopy, streamlining maintenance and reducing costs. Additionally, irrigation decisions based on farmers' empirical knowledge can be sufficiently made from above ground, leveraging their visual assessment of the canopy. As an alternative to visual inspection, image-based methods have been explored. Matsumura et al. (2006) demonstrated the feasibility of estimating leaf water stress by processing thermal infrared images of grape and satsuma leaves. Similarly, Ballester et al. ( 2013 ) reported on evaluating water stress from canopy temperature using thermal infrared images of persimmon and citrus, employing the Crop Water Stress Index (CWSI). Gonzalez-Dugo et al. ( 2013 ) utilized thermal drones to map species-specific water stress in orchards cultivating five varieties of fruit trees. However, these methods necessitate calculating canopy temperature to estimate water stress, requiring drones or thermal cameras, thereby complicating remote and automatic evaluation. Hence, research employing Convolutional Neural Networks (CNN), a deep learning method adept at image classification, has garnered considerable attention. CNN efficiently extracts and classifies complex patterns and features within images, making it widely applicable across various domains. For instance, in the medical field, CNN achieves over 90% accuracy in detecting various diseases like stroke, lung diseases, and hypertension (Liu et al., 2019 ), and it has been anticipated as a support tool for less experienced technicians in detecting variations such as cracks in concrete structures with around 90% accuracy (Aoshima et al., 2018 ). In agriculture, studies have demonstrated over 90% classification accuracy in identifying five diseases occurring on tomato leaves (Sarma et al., 2022 ) and in classifying ten classes of diseases on citrus fruits and leaves (Shermila et al., 2024 ), underscoring CNN's potential as a high-precision tool usable by inexperienced technicians when skilled expertise is required. This study endeavors to develop and assess a model for decision-making of persimmon irrigation utilizing CNN from canopy images. The model consists of two components: a soil moisture estimation model to gauge soil moisture levels and an irrigation decision-making model designed to ascertain the need for irrigation, drawing upon the irrigation criteria outlined by Kamoda ( 1987 ). 2. Materials and Method 2 − 1. Data for Model Construction Data collection took place in persimmon orchards situated in the Tochibara I and Hirahara districts within the Gose-Yoshino region of Nara Prefecture, Japan. These orchards were established as part of a national initiative and are equipped with sprinkler irrigation systems operational from June 1st to September 30th annually. For this study, Images and corresponding soil moisture data were collected during the irrigation seasons of two consecutive years: from June 30th to September 30th, 2022, and from June 1st to September 30th, 2023, between 6:00 and 18:00. The images captured by cameras are depicted in Fig. 3, with those obtained in 2022 labeled as "Flame1" and "Flame2," while those from 2023 are labeled as "Flame3" and "Flame4." Image data were acquired utilizing two fixed-point cameras (SCURA DVR-T1) positioned at each orchard, capturing images on an hourly basis. Soil moisture data were logged using METER's EC-5 soil moisture sensors, strategically buried at a depth of 25 cm beneath the surface. These sensors measure volumetric water content, which is subsequently translated into pF values through the outcomes of water retention tests (refer to Fig. 4a and Fig. 4b). To facilitate soil moisture assessment and irrigation determinations, we assumed water movement from roots to leaves, utilizing the soil moisture reading from 5 hours prior to the image capture time as the reference for each image. 2–2. Model Development CNNs, or Convolutional Neural Networks, represent a ubiquitous architecture in deep learning, comprising convolutional layers, pooling layers, and fully connected layers. Convolutional layers apply filters, serving as weights, to input data to extract local image patterns. Subsequently, pooling layers summarize and reduce the size of features gleaned by convolutional layers. Through the operations of convolutional and pooling layers, the extraction of image features is facilitated. The fully connected layer then outputs the extracted features as the final classification results. Furthermore, in this study, a dropout layer was incorporated to mitigate overfitting. This layer prevents certain neurons in the intermediate layers from being activated during training with a specified probability. The model employed in this study is outlined in Fig. 5 . It comprises a CNN architecture featuring an input layer, three convolutional blocks, a fully connected layer, and an output layer. Each block encompasses two convolutional layers, one pooling layer, and one dropout layer. The output items were classified into three categories for the soil moisture estimation model and two categories for the irrigation decision model. Further specifics of the models are elucidated below. 2–3. Soil Moisture Estimation Model The pF values' range divided into three ranges in accordance with soil moisture levels,. The classifications based on the pF values for the Tochihara I district were TD1 for values equal to or greater than 2.36, TD2 for values greater than or equal to 1.18 but less than 2.36, and TD3 for values less than 1.18. Similarly, for the Heibara district, the classifications were HD1 for values equal to or greater than 2.73, HD2 for values greater than or equal to 2.25 but less than 2.73, and HD3 for values less than 2.25. TD1 and HD1 signify a tendency towards dryness, while TD3 and HD3 indicate a tendency towards moisture. Stationary camera images were classified into each category based on the corresponding soil moisture content and were utilized for training. The total number of image data used for training was 4,198 for the Tochihara I and 5,568 for the Heibara. These datasets were partitioned for training and evaluation as illustrated in Table 1 . The resultant models were denoted as TSm for the Tochihara I and HSm for the Heibara. Following this, the acquired models were utilized to categorize the verification data into three distinct classes and subsequently compared with the actual classifications. The comparison outcomes were organized into a confusion matrix as depicted in Table 2 . A confusion matrix effectively summarizes the classification results of classes, facilitating the determination of both correct and incorrect classifications. The evaluation of the models was conducted using accuracy (Eq. (2)) and F1-scores for each class (Equations (3)-(5)). In Eq. (3), class1 denotes TD1 and HD1, class2 represents TD2 and HD2, while class3 corresponds to TD3 and HD3. The F1-score represents the harmonic mean of two key metrics: recall, which signifies the proportion of correct predictions out of the total correct data, and precision, which denotes the proportion of correct predictions out of all predictions made. With a scale from 0 to 1, higher F1-score values reflect greater accuracy. $$Accuracy= \frac{a+e +i}{a +b +c +d +e+f+g+h+i} \left(2\right)$$ $$F1-score-class1= \frac{2a}{2a +b +c +d +g} \left(3\right)$$ $$F1-score-class2= \frac{2e}{2e +b +h +a +c} \left(4\right)$$ $$F1-score-class3= \frac{2i}{2i +c +f +h +g} \left(5\right)$$ 2–4. Irrigation Decision Model Following Kamoda's (1987) recommendations for soil moisture levels, the image data were classified into two categories: "irrigation" if the soil moisture fell below the recommended level and "appropriate" if sufficient moisture was present. Kamoda's threshold for soil dryness varies across fruit growth stages, which are divided into three periods: early growth (April-May), fruit enlargement (June-August), and ripening (September-October). With the threshold altering from June to August (pF value equal to or greater than 2.7) and from September onward (pF value equal to or greater than 3.0), adjustments were made accordingly for classification during these periods. Training and evaluation proceeded based on the distribution of images as outlined in Table 3 . The models derived were designated as TIr model for the Tochihara I, HIr model for the Heibara, and CIr model for a combined model trained on both districts, aiming to explore the potential for constructing a unified model applicable across multiple districts. Subsequently, these models were employed to classify the test data for irrigation necessity and juxtaposed with the actual classifications. The comparative results were structured using a confusion matrix depicted in Table 4 , followed by evaluation utilizing Equations (6)-(8) as delineated below. $$Accuracy= \frac{a+d}{a +b +c +d } \left(6\right)$$ $$F1-score-Irrigation= \frac{2a}{2a +c +d} \left(7\right)$$ $$F1-score-appropriate= \frac{2d}{2d +c +b} \left(8\right)$$ 3. Result and Discussion 3 − 1. Soil Moisture Table 5 displays the verification outcomes of the TSm model, exhibiting an accuracy of 0.88. The F1-scores for TD1, TD2, and TD3 were 0.94, 0.72, and 0.75, respectively. Conversely, Table 6 presents the verification results of the HSm model, showcasing an accuracy of 0.91. The F1-scores for HD1, HD2, and HD3 were 0.94, 0.85, and 0.72, respectively. Both models demonstrated sufficient accuracy, with precision hovering around 0.9 concerning accuracy and F-scores. Nonetheless, variations in accuracy were discernible due to dissimilarities in the number of training images for each class. Regarding misclassification counts for each class in the Tochihara I district, there were 27 misclassifications for TD1, with the majority erroneously classified as TD2. Furthermore, there were 22 misclassifications for TD3, with the majority inaccurately classified as TD2. Similarly, in the Heibara district, misclassifications of HD2 were predominant in both HD1 and HD3 classes. These findings imply that distinguishing subtle differences in soil moisture levels, such as between TD1 and TD2 or HD1 and HD2, or between TD2 and TD3 or HD2 and HD3, poses a challenge, while classification based on significant differences in soil moisture, such as between TD1 and TD3 or HD1 and HD3, is more feasible. However, considering the potential impact of the uneven distribution of training images for each class, further data accumulation and verification are imperative for future enhancements. 3 − 2. Irrigation Index Table 7 illustrates the verification outcomes of the TIr model, indicating an accuracy of 0.93. The F1-scores for "irrigation" and "appropriate" were 0.93 and 0.94, respectively. Meanwhile, Table 8 showcases the verification results of the HIr model, with an accuracy of 0.88. The F1-scores for "irrigation" and "appropriate" were 0.87 and 0.88, respectively. Both districts demonstrated high accuracy, around 0.90, for "irrigation" and "appropriate" classifications, underscoring the feasibility of accurately discerning the necessity of irrigation. To explore the viability of constructing a unified model for the each district, a model trained on combined images from both regions was developed, with results detailed in Table 9 . The achieved accuracy was 0.71, with F1-scores for "irrigation" and "appropriate" classifications at 0.74 and 0.68, respectively, lower than those observed for each residential area individually. Furthermore, the verification outcomes of the Heibara district images using the TIr model (Table 10 ) and the Tochihara I district images using the HIr model (Table 11 ) are provided below. Verification accuracy and overall F1-score for the Heibara district using the TIr model were 0.50 and 0.47, respectively. Similarly, verification results for the Tochihara I district using the HIr model yielded an accuracy and overall F1-score of 0.51. Both demonstrated lower accuracy compared to individual district verifications, suggesting that models tailored to each region were developed due to training solely on images from their respective areas. Moreover, the construction and evaluation of a joint model with a limited number of districts may have failed to capture common features among persimmon trees. Consequently, the concept of building an irrigation decision model adaptable to various fields by acquiring and training images from multiple fields warrants further investigation. Thus, future endeavors should delve into the feasibility of this approach, particularly concerning the number of fields involved. Table 12 presents the proportion of misclassified images from each camera installed in the respective districts. Notably, for Flame 4 in the Tochihara I district and Flame 2 and Flame 4 in the Heibara district, which captured the entire tree, the proportions were notably high, at 35%, 43%, and 27%, respectively. As persimmon trees mature, their branches gradually bend under the weight of the fruit. Consequently, in Flame 2 of the Heibara district, initially, roughly half of the image was occupied by leaves. However, as time progressed, about half of the image was dominated by the ground. This resulted in a decrease in the proportion of leaves in the image, leading to an increase in misclassifications. Conversely, Flame 1 and Flame 3 in the Heibara district, which closely captured the leaves, exhibited lower proportions of misclassified images, at 13% and 17%, respectively, compared to images capturing the entire tree. Hence, it is preferable to capture images from locations where the angle of the leaves in the image remains consistent over time, facilitating maintenance. Furthermore, images capturing the entire tree had a lower proportion of misclassifications when leaves occupied a larger proportion of the image, suggesting that classification is easier when leaves dominate the image. Additionally, misclassified images occasionally featured unclear features due to backlighting or water droplets on the camera after rainfall. While these images were included in the training dataset, excluding them from training and evaluation is anticipated to enhance classification accuracy. Figure 6 and 7 depict the decision criteria of the models following the methodology outlined by Zhou et al. ( 2016 ) and Selvaraju et al. ( 2017 ). The color bar in the figures represents normalized feature weights, with red indicating higher importance for classification and blue indicating lower contribution. Notably, in images capturing the entire tree in both districts, red heatmaps are predominantly concentrated on leaves closer to the camera, while blue heatmaps, indicating lower importance, are more prevalent for distant leaves. Furthermore, for images capturing leaves closely in the Heibara district, the heatmaps tend to be concentrated on the side of the leaves. This observation aligns with the common knowledge that leaves tend to wilt or curl when experiencing water stress, suggesting that the models may be leveraging this physiological characteristic to ascertain the necessity of irrigation. In summary, to enhance the effectiveness of the models, it is essential to install cameras closer to the leaves during image capture and utilize installation methods capable of capturing changes in leaf shape. 4. Conclusion In this study, by constructing models to estimate soil moisture from leaf images and to determine the necessity of irrigation, we were able to establish a correlation between leaf images and soil moisture levels, suggesting the potential for new irrigation indicators. While the accuracy of the soil moisture estimation model was likely affected by imbalances in the number of samples for each class, a highly accurate model was constructed for classes with a sufficient number of samples. Previous studies have reported that increasing irrigation improved fruit growth, but excessive soil moisture led to increased physiological fruit drop (Matsuda et al., 2013 ). Additionally, relationships have been found between irrigation levels and fruit size (Matsumura, 1999 ), as well as between irregular fruit shapes and the proportion of unsuitable fruit for shipping (Kanety et al., 2014 ). Therefore, precise irrigation management is necessary to improve quality and increase yield. Consequently, more detailed estimation of soil moisture is crucial. In the irrigation decision model, high-accuracy models were constructed for eachdistrict. However, the CIr model, blending images from two fields, exhibited low accuracy. Additionally, verification results from the Heibara district using the TIr model and the Tochihara I district using the HIr model displayed low accuracy, suggesting the development of specialized models for each field, possibly due to incomplete understanding of persimmon's general characteristics. Consequently, future endeavors should prioritize constructing versatile models that do not specialize in each field by augmenting the number of training samples and quantitatively evaluating the models' decision criteria. This model holds promise for facilitating remote irrigation decisions by integrating field-acquired images into edge devices for on-site processing. Moreover, when combined with remote irrigation systems currently in development, it can pave the way for automated irrigation, offering significant potential for labor-saving measures. Declarations Author Contribution A.O. wrote the main manuscript text and program coding, Data curation. A.Y. Data curation, M.K. Conceptualization, Methodology development, Y.M. Conceptualization, Methodology development, Research organization, Editing. All authors reviewed manuscript. Acknowledgement This research received support from the development and improvement program of strategic smart agricultural technology grants provided by the Project of the Bio-oriented Technology Research Advancement Institution (BRAIN). The authors extend their sincere appreciation to Nara Prefecture for their support and facilitation of the research activities. References Aoshima, W., Kawamura, S., Nakano, S., & Nakamura, H. (2018). Study on Variant Extraction of concrete structures using Image Recognition by Deep Learning. The Journal of the Japan Society of Civil Engineers E2 (Materials and Concrete Structures), 74(4) , 293-305. (In Japanese) https://www.jstage.jst.go.jp/article/jscejmcs/74/4/74_293/_pdf/-char/ja Ballester, C., Jiménez-Bello, M. A., Castel, J. R., & Intrigliolo, D. S. (2013). Usefulness of thermography for plant water stress detection in citrus and persimmon trees. Agricultural and forest Meteorology, 168 , 120-129. https://www.sciencedirect.com/science/article/pii/S0168192312002572 Gojo-Yoshino Land Improvement District. Leaflet of Gojo-Yoshino Land Improvement District; Gojo-Yoshino Land Improvement District: Nara, Japan, 2002. (In Japanese) Gonzalez-Dugo, V., Zarco-Tejada, P., Nicolás, E., Nortes, P. A., Alarcón, J. J., Intrigliolo, D. S., & Fereres, E. J. P. A. (2013). Using high resolution UAV thermal imagery to assess the variability in the water status of five fruit tree species within a commercial orchard. Precision Agriculture, 14 , 660-678. https://link.springer.com/article/10.1007/s11119-013-9322-9 Iwata, Y., Nakagawa, F., Aida, N., Nawa, N., Miyamoto, T., Kameyama, K., & Syoubu, J. (2019). Effect of Drip Irrigation on the Weight of Fruits of Japanese Persimmon (Diospyros kaki Thunb.) at the Field in Sado Island, Japan. Transactions of the Japanese Society of Irrigation, Drainage and Rural Engineering, 87(2) , I_227 - I_237. (In Japanese) https://www.jstage.jst.go.jp/article/jsidre/87/2/87_I_227/_article/-char/ja/ Japan Meteorological Agency (JMA). (2023). https://www.data.jma.go.jp/obd/stats/etrn/index.php (in Japanese). Kamoda, S. (1987): Environmental Control in Facility Cultivation of Fruit Trees. Abstracts of Symposium of the Horticultural Society , 1-8. (In Japanese) Kanety. T., Naor A., Dicken. U., Lemcoff H., Cohen. S. (2014). Irrigation influences on growth, yield, and water use of persimmon trees. Irrigation Science, 32 , 1-13. (In Japanese)https://link.springer.com/article/10.1007/s00271-013-0408-y Tanaka, K., & Aoki, M. (1971). Effect of Irrigation and Nitrogen Application in Summer on the Fruiting of ‘Fuyu’ Kaki. Research bulletin of the Aichi-ken Agricultural Research Center. Series B, 3 , 9-18. (In Japanese) Liu. W., Qin. C., Gao. K., Li. H., Qin. Z., Cao. Y., Si. Wen. (2019). Research on Medical Data Feature Extraction and Intelligent Recognition Technology Based on Convolutional Neural Network. IEEE Access, 7 , 150157 – 150167. https://ieeexplore.ieee.org/document/8849989 Matsuda, M., Habu, T., Konishi, T., Nonaka, K., Kusumi, K., Kurosawa, T., & Kitajima, A. (2011). Effect of the drip irrigation system with plastic mulch on fruit quality in Japanese persimmon cv. ‘Taisyu’. Bulletin of the Experimental Farm, Kyoto University, 20 , 29-32. (In Japanese) https://agriknowledge.affrc.go.jp/RN/2030830743 Matsuda, M., Matsumoto, D., Konishi, T., Nonaka, K., Kusumi, K., Kurosawa, T., & Kitajima, A. (2013). Effect of the drip irrigation system with plastic mulch on fruit quality in Japanese persimmon cv. ‘Taisyu’. Ⅱ. Bulletin of the Experimental Farm, Kyoto University, 22 , pp27-32. (In Japanese) https://agriknowledge.affrc.go.jp/RN/2030872756 Matsumura, H. (1999). Container Planting Cultivation of Japanese Persimmon (Diospyros kaki Thunb.) (Pat3) -The Improvement of the Fruit Quality by Fertilization and irrigation. Journal of Society of High Technology in Agricultur e , 11(4) , 259-266. (In Japanese) https://www.jstage.jst.go.jp/article/jshita1991/11/4/11_4_259/_article/-char/ja/ Muramatsu, N., & Hiraoka, K. (2006). Estimation of Water Content in the Leaves of Fruit Trees Using Infra-red Images. Horticultural Research, 5(4) , 397-402. (In Japanese) https://www.jstage.jst.go.jp/article/hrj/5/4/5_4_397/_article/-char/ja/ Okatani, T. (2022). Deep Learning, Revised 2nd Edition. Kodansha . (In Japanese) O'Shea, K., & Nash, R. (2015). An introduction to convolutional neural networks. arXiv preprint. (In Japanese) https://arxiv.org/abs/1511.08458 Sarma, K. K., Das, K. K., Mishra, V., Bhuiya, S., & Kaplun, D. (2022). Learning Aided System for Agriculture Monitoring Designed Using Image Processing and IoT-CNN. IEEE Access, 10 , 41525-41536. https://ieeexplore.ieee.org/document/9756536 Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, 618-626. https://arxiv.org/abs/1610.02391 Shermila, P. J., Victor, A., Manoj, S. O., & Devi, E. A. (2024). Automatic detection and classification of disease in citrus fruit and leaves using a customized CNN based model. Boletín Latinoamericano y del Caribe de Plantas Medicinales y Aromáticas, 23(2) , 180-198. https://www.blacpma.ms-editions.cl/index.php/blacpma/article/view/409/408 Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., & Torralba, A. (2016). Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition , 2921-2929. https://arxiv.org/abs/1512.04150 Tables Tables 1 to 12 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Tables.docx Cite Share Download PDF Status: Published Journal Publication published 14 Jan, 2025 Read the published version in Environmental Monitoring and Assessment → Version 1 posted Editorial decision: Revision requested 06 Jun, 2024 Editor assigned by journal 06 Jun, 2024 Submission checks completed at journal 06 Jun, 2024 First submitted to journal 11 May, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4404121","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":309394730,"identity":"9940516b-29f7-417a-ab50-d3695fc2cdd5","order_by":0,"name":"Atsushi Okayama","email":"","orcid":"","institution":"Kindai University","correspondingAuthor":false,"prefix":"","firstName":"Atsushi","middleName":"","lastName":"Okayama","suffix":""},{"id":309394731,"identity":"81cb9acc-4b07-4e0f-aaa1-0b8f951104a8","order_by":1,"name":"Atsushi Yamamoto","email":"","orcid":"","institution":"Kindai University","correspondingAuthor":false,"prefix":"","firstName":"Atsushi","middleName":"","lastName":"Yamamoto","suffix":""},{"id":309394732,"identity":"b01956d1-ed86-411a-9703-5d8beb362daa","order_by":2,"name":"Masaomi Kimura","email":"","orcid":"","institution":"Kindai University","correspondingAuthor":false,"prefix":"","firstName":"Masaomi","middleName":"","lastName":"Kimura","suffix":""},{"id":309394733,"identity":"01f93622-8bc3-40c1-9c55-2c182c66e662","order_by":3,"name":"Yutaka Matsuno","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCUlEQVRIiWNgGAWjYBADOQh1QAKIGRtATGbciiFSxqRrSWyAaIFifEC3/fzhz7w7bNK3s599+IHhjEUe3/HDDQw/ahjYzXFoMTuTzCbNeyYtd2dPurEEww2JYskziQ2MPccYmC0bcGg5kMzGzNt2OHfDgTQGCYYPEokbDgAdydvAwGyAw4Vm5x8zfwZqSTc4/4z5B1jL+YcNjH/xabmRzCAN1JJgcCONDeSwxA03EhuY8dpy47GZ5Ny2NMMNN56xWSSckUiceeNhw2GZYxK4/XI+8fGHt2028gbn05hvfDhWl9h3Pv3hwzc1Nsm4QgwVJEBpoJMkkg2I0oIM7EjXMgpGwSgYBcMUAAC6dGCHraAdaAAAAABJRU5ErkJggg==","orcid":"","institution":"Kindai University","correspondingAuthor":true,"prefix":"","firstName":"Yutaka","middleName":"","lastName":"Matsuno","suffix":""}],"badges":[],"createdAt":"2024-05-11 07:33:46","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4404121/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4404121/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s10661-024-13602-1","type":"published","date":"2025-01-14T15:57:01+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":57627087,"identity":"f01a094b-92a2-4761-ad45-3b20c09d80e3","added_by":"auto","created_at":"2024-06-03 14:11:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":870379,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe study area (Gojo Yoshino area of Nara Prefecture).\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a) Location map in Japan; (b) open field orchards widely grown in the mountainous region; (c) Greenhouse persimmon cultivation areas; (d) Ichinoki Dam for irrigation supply.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/e5cb8b91c6da491576aabd75.png"},{"id":57625476,"identity":"091336d1-7562-4d9d-a37e-37b28e01c600","added_by":"auto","created_at":"2024-06-03 13:55:34","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":384222,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview Schematic of Research Project Utilizing New LPWAN Communication Standard\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/edc7dd71d3d06bbc4a783f57.png"},{"id":57626431,"identity":"5fff00b8-0cbb-43eb-ac88-556cf83335e9","added_by":"auto","created_at":"2024-06-03 14:03:34","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1496237,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExamples of images from different cameras in Tochihara I District (a) and Hirabara District (b). \u003c/strong\u003e\u0026nbsp;\u003cstrong\u003eFlame1 and Flame2 were captured in 2022、Flame3 and Flame4 were captured in 2022.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/126ab9e0f7f0adc83aa3170a.png"},{"id":57626430,"identity":"a0cb5c87-1f44-4cce-86ed-3b813c79c614","added_by":"auto","created_at":"2024-06-03 14:03:34","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":48121,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe moisture characteristic curves of Tochihara Ⅰ District (a) \u003c/strong\u003e\u0026nbsp;\u003cstrong\u003eand Heibara District (b)\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/aa52f611fd2cf3142a4fa074.png"},{"id":57625480,"identity":"24a74a5c-b62b-41e9-ae20-88b359036022","added_by":"auto","created_at":"2024-06-03 13:55:34","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":84873,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview diagram of the constructed model\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/69f67206d9e6998c8c01ced1.png"},{"id":57625482,"identity":"b5a6676e-01ec-4f5d-a50c-60fa46d4ce97","added_by":"auto","created_at":"2024-06-03 13:55:34","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":901941,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualization images depicting judgments made by the TIr model for each Flame. The left corresponds to 'irrigation', while the right corresponds to 'appropriate'.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/0e0a0d3e339e3cbca6d186b1.png"},{"id":57625483,"identity":"2147a343-c65c-4da2-a6c9-23bb35fd2542","added_by":"auto","created_at":"2024-06-03 13:55:34","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":892465,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualization images depicting judgments made by the HIr model for each Flame. The left corresponds to 'irrigation', while the right corresponds to 'appropriate'.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/09abfeabebf92877efbd30d6.png"},{"id":74284506,"identity":"245fcb90-230f-4b2c-a499-b91868e93a3d","added_by":"auto","created_at":"2025-01-20 16:08:05","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5619350,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/90eef177-da96-4064-8a89-aeda9fd8be7e.pdf"},{"id":57625478,"identity":"2b9b58ee-b709-4971-b39d-75aaa500c5d6","added_by":"auto","created_at":"2024-06-03 13:55:34","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":720956,"visible":true,"origin":"","legend":"","description":"","filename":"Tables.docx","url":"https://assets-eu.researchsquare.com/files/rs-4404121/v1/14d0bf5ad8c2f062d0618972.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Utilizing Convolutional Neural Networks (CNN) for Persimmon Irrigation Decision-Making: A Case Study in the Gojo Yoshino Region, Nara Prefecture, Japan","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eIn Japan, persimmons have held a prominent status among the nation\u0026apos;s fruits for centuries, with a cultivation history dating back to ancient times. The commercialization of persimmon cultivation can be traced to the 1920s in the Gojo Yoshino region of Nara Prefecture, Japan. Situated in the mountainous core of Japan (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e-a), this area has an average annual temperature of 15.3\u0026deg;C and annual precipitation of 1,492 mm, as recorded in 2023 by the Japan Meteorological Agency (JMA). The 1970s saw a significant increase in land development initiatives aimed at facilitating persimmon cultivation across the rugged terrain, driven by the inception of a national agricultural development project in 1974. Consequently, the Gojo Yoshino region emerged as Nara Prefecture\u0026apos;s primary persimmon-producing hub, eventually becoming Japan\u0026apos;s second-largest persimmon-producing prefecture. Cultivation methods in this region encompass both greenhouse and open-field practices (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e-b). To ensure consistent fruit yields, irrigation water is sourced from the Ichinoki Dam (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e-c), a national government undertaking completed in 1997. The dam supports an extensive irrigation network, covering a total area of 1,817 hectares, with approximately 1,600 hectares dedicated to persimmon cultivation (Gojo-Yoshino Land Improvement District, \u003cspan class=\"CitationRef\"\u003e2002\u003c/span\u003e).\u003c/p\u003e\n\u003cp\u003eIn persimmon cultivation, numerous studies have underscored the substantial influence of irrigation quantity on fruit size and yield. For instance, Aoki et al. (1971) conducted a study comparing different treatment zones involving irrigation and nitrogen application for \u0026apos;Fuyu\u0026apos; persimmon fruit. They observed that the irrigation treatment zone yielded more fruits, with a higher proportion falling into the L-grade or higher category compared to the non-irrigated treatment zone. Similarly, investigations on the multi-cultivation irrigation system for \u0026apos;Taiaki\u0026apos; persimmon fruit, carried out by Matsuda et al. (\u003cspan class=\"CitationRef\"\u003e2011\u003c/span\u003e), revealed that the irrigation treatment zone produced larger fruits than the non-treated zone. These findings strongly indicate that increased irrigation can significantly augment harvest yields.\u003c/p\u003e\n\u003cp\u003eIn the Gojo Yoshino region, irrigation of open fields is typically scheduled during the summer months, spanning from June to September, coinciding with the growth period of persimmon fruits. Each field is assigned specific irrigation times, typically occurring for one hour every three days. To accommodate farmers\u0026apos; schedules and mitigate the impact of peak working hours, irrigation is often timed for early mornings or late evenings extending to midnight. However, due to the scattered distribution of fields across steep mountain slopes, accessing these sites demands significant effort and time from farmers.\u003c/p\u003e\n\u003cp\u003eIn an effort to streamline this labor-intensive process, plans are underway to implement a low-power wide-area network (LPWAN) featuring an automatic and remote irrigation system (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). Distinguishing itself from conventional LPWANs such as LoRa, this relatively new system offers an impressive communication range of 100 km on flat terrain, representing a tenfold improvement. Initial tests of a prototype system have already demonstrated successful communication across a mountaintop spanning approximately 10 km.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCommunication Standard\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCurrently, persimmon irrigation decision-making relies on the experience and judgment of individual growers who observe the trees in the field. Even with the introduction of the new LPWAN system, which enables remote control and automation of irrigation, the workload of farmers may not decrease if they still need to physically visit the fields to make irrigation decisions. Therefore, there is a pressing need for technology that facilitates remote irrigation decision-making.\u003c/p\u003e\n\u003cp\u003eSoil moisture monitoring stands as a well-established technique for irrigation scheduling. However, the variability of soil moisture across different soil types and locations presents a challenge to its practical application. To address this challenge, the assessment of soil water potential, quantified by the pF value, has been proposed. This metric enables the determination of water holding capacity across diverse soil types and facilitates tailored irrigation scheduling to meet crop-specific requirements. The pF value is defined as the logarithm of the negative pressure (measured in cm of water) necessary to extract water from the soil, as depicted in Eq.\u0026nbsp;(1):\u003c/p\u003e\n\u003cp\u003epF\u0026thinsp;=\u0026thinsp;log10(\u0026minus;\u0026thinsp;h) (1)\u003c/p\u003e\n\u003cp\u003eWhere h represents the pressure head in cm of water.\u003c/p\u003e\n\u003cp\u003eUtilizing the pF value as a benchmark for persimmon irrigation, Kamoda (\u003cspan class=\"CitationRef\"\u003e1987\u003c/span\u003e) proposed optimal pF values tailored to each growth stage. Furthermore, studies, such as that by Iwata et al. (\u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e), indicate the effectiveness of soil moisture management below pF 3.0, a threshold considered growth inhibitory. While remote irrigation decision-making based on these values is feasible, maintenance poses challenges due to the necessity of burying sensors, rendering visual inspection of equipment impossible. Moreover, this method incurs costs associated with sensor re-burial in case of malfunctions or calibration. Consequently, attention has shifted towards irrigation decision-making grounded in canopy inspection from above ground. Unlike buried sensors, this approach targets the visible canopy, streamlining maintenance and reducing costs. Additionally, irrigation decisions based on farmers\u0026apos; empirical knowledge can be sufficiently made from above ground, leveraging their visual assessment of the canopy. As an alternative to visual inspection, image-based methods have been explored. Matsumura et al. (2006) demonstrated the feasibility of estimating leaf water stress by processing thermal infrared images of grape and satsuma leaves. Similarly, Ballester et al. (\u003cspan class=\"CitationRef\"\u003e2013\u003c/span\u003e) reported on evaluating water stress from canopy temperature using thermal infrared images of persimmon and citrus, employing the Crop Water Stress Index (CWSI). Gonzalez-Dugo et al. (\u003cspan class=\"CitationRef\"\u003e2013\u003c/span\u003e) utilized thermal drones to map species-specific water stress in orchards cultivating five varieties of fruit trees. However, these methods necessitate calculating canopy temperature to estimate water stress, requiring drones or thermal cameras, thereby complicating remote and automatic evaluation. Hence, research employing Convolutional Neural Networks (CNN), a deep learning method adept at image classification, has garnered considerable attention. CNN efficiently extracts and classifies complex patterns and features within images, making it widely applicable across various domains. For instance, in the medical field, CNN achieves over 90% accuracy in detecting various diseases like stroke, lung diseases, and hypertension (Liu et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e), and it has been anticipated as a support tool for less experienced technicians in detecting variations such as cracks in concrete structures with around 90% accuracy (Aoshima et al., \u003cspan class=\"CitationRef\"\u003e2018\u003c/span\u003e). In agriculture, studies have demonstrated over 90% classification accuracy in identifying five diseases occurring on tomato leaves (Sarma et al., \u003cspan class=\"CitationRef\"\u003e2022\u003c/span\u003e) and in classifying ten classes of diseases on citrus fruits and leaves (Shermila et al., \u003cspan class=\"CitationRef\"\u003e2024\u003c/span\u003e), underscoring CNN\u0026apos;s potential as a high-precision tool usable by inexperienced technicians when skilled expertise is required.\u003c/p\u003e\n\u003cp\u003eThis study endeavors to develop and assess a model for decision-making of persimmon irrigation utilizing CNN from canopy images. The model consists of two components: a soil moisture estimation model to gauge soil moisture levels and an irrigation decision-making model designed to ascertain the need for irrigation, drawing upon the irrigation criteria outlined by Kamoda (\u003cspan class=\"CitationRef\"\u003e1987\u003c/span\u003e).\u003c/p\u003e"},{"header":"2. Materials and Method","content":"\u003ch3\u003e2\u0026thinsp;\u0026minus;\u0026thinsp;1. Data for Model Construction\u003c/h3\u003e\n\u003cp\u003eData collection took place in persimmon orchards situated in the Tochibara I and Hirahara districts within the Gose-Yoshino region of Nara Prefecture, Japan. These orchards were established as part of a national initiative and are equipped with sprinkler irrigation systems operational from June 1st to September 30th annually. For this study, Images and corresponding soil moisture data were collected during the irrigation seasons of two consecutive years: from June 30th to September 30th, 2022, and from June 1st to September 30th, 2023, between 6:00 and 18:00. The images captured by cameras are depicted in Fig. 3, with those obtained in 2022 labeled as \u0026quot;Flame1\u0026quot; and \u0026quot;Flame2,\u0026quot; while those from 2023 are labeled as \u0026quot;Flame3\u0026quot; and \u0026quot;Flame4.\u0026quot;\u003c/p\u003e\n\u003cp\u003eImage data were acquired utilizing two fixed-point cameras (SCURA DVR-T1) positioned at each orchard, capturing images on an hourly basis. Soil moisture data were logged using METER\u0026apos;s EC-5 soil moisture sensors, strategically buried at a depth of 25 cm beneath the surface. These sensors measure volumetric water content, which is subsequently translated into pF values through the outcomes of water retention tests (refer to Fig. 4a and Fig. 4b). To facilitate soil moisture assessment and irrigation determinations, we assumed water movement from roots to leaves, utilizing the soil moisture reading from 5 hours prior to the image capture time as the reference for each image.\u003c/p\u003e\n\u003ch3\u003e2\u0026ndash;2. Model Development\u003c/h3\u003e\n\u003cp\u003eCNNs, or Convolutional Neural Networks, represent a ubiquitous architecture in deep learning, comprising convolutional layers, pooling layers, and fully connected layers. Convolutional layers apply filters, serving as weights, to input data to extract local image patterns. Subsequently, pooling layers summarize and reduce the size of features gleaned by convolutional layers. Through the operations of convolutional and pooling layers, the extraction of image features is facilitated. The fully connected layer then outputs the extracted features as the final classification results. Furthermore, in this study, a dropout layer was incorporated to mitigate overfitting. This layer prevents certain neurons in the intermediate layers from being activated during training with a specified probability.\u003c/p\u003e\n\u003cp\u003eThe model employed in this study is outlined in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. It comprises a CNN architecture featuring an input layer, three convolutional blocks, a fully connected layer, and an output layer. Each block encompasses two convolutional layers, one pooling layer, and one dropout layer. The output items were classified into three categories for the soil moisture estimation model and two categories for the irrigation decision model. Further specifics of the models are elucidated below.\u003c/p\u003e\n\u003ch3\u003e2\u0026ndash;3. Soil Moisture Estimation Model\u003c/h3\u003e\n\u003cp\u003eThe pF values\u0026apos; range divided into three ranges in accordance with soil moisture levels,. The classifications based on the pF values for the Tochihara I district were TD1 for values equal to or greater than 2.36, TD2 for values greater than or equal to 1.18 but less than 2.36, and TD3 for values less than 1.18. Similarly, for the Heibara district, the classifications were HD1 for values equal to or greater than 2.73, HD2 for values greater than or equal to 2.25 but less than 2.73, and HD3 for values less than 2.25. TD1 and HD1 signify a tendency towards dryness, while TD3 and HD3 indicate a tendency towards moisture. Stationary camera images were classified into each category based on the corresponding soil moisture content and were utilized for training. The total number of image data used for training was 4,198 for the Tochihara I and 5,568 for the Heibara. These datasets were partitioned for training and evaluation as illustrated in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. The resultant models were denoted as TSm for the Tochihara I and HSm for the Heibara.\u003c/p\u003e\n\u003cp\u003eFollowing this, the acquired models were utilized to categorize the verification data into three distinct classes and subsequently compared with the actual classifications. The comparison outcomes were organized into a confusion matrix as depicted in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. A confusion matrix effectively summarizes the classification results of classes, facilitating the determination of both correct and incorrect classifications. The evaluation of the models was conducted using accuracy (Eq. (2)) and F1-scores for each class (Equations (3)-(5)). In Eq. (3), class1 denotes TD1 and HD1, class2 represents TD2 and HD2, while class3 corresponds to TD3 and HD3.\u003c/p\u003e\n\u003cp\u003eThe F1-score represents the harmonic mean of two key metrics: recall, which signifies the proportion of correct predictions out of the total correct data, and precision, which denotes the proportion of correct predictions out of all predictions made. With a scale from 0 to 1, higher F1-score values reflect greater accuracy.\u003c/p\u003e\n\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$Accuracy= \\frac{a+e +i}{a +b +c +d +e+f+g+h+i} \\left(2\\right)$$\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e$$F1-score-class1= \\frac{2a}{2a +b +c +d +g} \\left(3\\right)$$\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e$$F1-score-class2= \\frac{2e}{2e +b +h +a +c} \\left(4\\right)$$\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e$$F1-score-class3= \\frac{2i}{2i +c +f +h +g} \\left(5\\right)$$\u003c/div\u003e\n\u003c/div\u003e\n\u003ch3\u003e2\u0026ndash;4. Irrigation Decision Model\u003c/h3\u003e\n\u003cp\u003eFollowing Kamoda\u0026apos;s (1987) recommendations for soil moisture levels, the image data were classified into two categories: \u0026quot;irrigation\u0026quot; if the soil moisture fell below the recommended level and \u0026quot;appropriate\u0026quot; if sufficient moisture was present. Kamoda\u0026apos;s threshold for soil dryness varies across fruit growth stages, which are divided into three periods: early growth (April-May), fruit enlargement (June-August), and ripening (September-October). With the threshold altering from June to August (pF value equal to or greater than 2.7) and from September onward (pF value equal to or greater than 3.0), adjustments were made accordingly for classification during these periods. Training and evaluation proceeded based on the distribution of images as outlined in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n\u003cp\u003eThe models derived were designated as TIr model for the Tochihara I, HIr model for the Heibara, and CIr model for a combined model trained on both districts, aiming to explore the potential for constructing a unified model applicable across multiple districts. Subsequently, these models were employed to classify the test data for irrigation necessity and juxtaposed with the actual classifications. The comparative results were structured using a confusion matrix depicted in Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, followed by evaluation utilizing Equations (6)-(8) as delineated below.\u003c/p\u003e\n\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e$$Accuracy= \\frac{a+d}{a +b +c +d } \\left(6\\right)$$\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e$$F1-score-Irrigation= \\frac{2a}{2a +c +d} \\left(7\\right)$$\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Equg\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equg\" name=\"EquationSource\"\u003e$$F1-score-appropriate= \\frac{2d}{2d +c +b} \\left(8\\right)$$\u003c/div\u003e\n\u003c/div\u003e"},{"header":"3. Result and Discussion","content":"\u003ch3\u003e3\u0026thinsp;\u0026minus;\u0026thinsp;1. Soil Moisture\u003c/h3\u003e\n\u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e displays the verification outcomes of the TSm model, exhibiting an accuracy of 0.88. The F1-scores for TD1, TD2, and TD3 were 0.94, 0.72, and 0.75, respectively. Conversely, Table \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e presents the verification results of the HSm model, showcasing an accuracy of 0.91. The F1-scores for HD1, HD2, and HD3 were 0.94, 0.85, and 0.72, respectively. Both models demonstrated sufficient accuracy, with precision hovering around 0.9 concerning accuracy and F-scores. Nonetheless, variations in accuracy were discernible due to dissimilarities in the number of training images for each class.\u003c/p\u003e\n\u003cp\u003eRegarding misclassification counts for each class in the Tochihara I district, there were 27 misclassifications for TD1, with the majority erroneously classified as TD2. Furthermore, there were 22 misclassifications for TD3, with the majority inaccurately classified as TD2. Similarly, in the Heibara district, misclassifications of HD2 were predominant in both HD1 and HD3 classes. These findings imply that distinguishing subtle differences in soil moisture levels, such as between TD1 and TD2 or HD1 and HD2, or between TD2 and TD3 or HD2 and HD3, poses a challenge, while classification based on significant differences in soil moisture, such as between TD1 and TD3 or HD1 and HD3, is more feasible. However, considering the potential impact of the uneven distribution of training images for each class, further data accumulation and verification are imperative for future enhancements.\u003c/p\u003e\n\u003ch3\u003e3\u0026thinsp;\u0026minus;\u0026thinsp;2. Irrigation Index\u003c/h3\u003e\n\u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e illustrates the verification outcomes of the TIr model, indicating an accuracy of 0.93. The F1-scores for \u0026quot;irrigation\u0026quot; and \u0026quot;appropriate\u0026quot; were 0.93 and 0.94, respectively. Meanwhile, Table \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e showcases the verification results of the HIr model, with an accuracy of 0.88. The F1-scores for \u0026quot;irrigation\u0026quot; and \u0026quot;appropriate\u0026quot; were 0.87 and 0.88, respectively. Both districts demonstrated high accuracy, around 0.90, for \u0026quot;irrigation\u0026quot; and \u0026quot;appropriate\u0026quot; classifications, underscoring the feasibility of accurately discerning the necessity of irrigation.\u003c/p\u003e\n\u003cp\u003eTo explore the viability of constructing a unified model for the each district, a model trained on combined images from both regions was developed, with results detailed in Table \u003cspan class=\"InternalRef\"\u003e9\u003c/span\u003e. The achieved accuracy was 0.71, with F1-scores for \u0026quot;irrigation\u0026quot; and \u0026quot;appropriate\u0026quot; classifications at 0.74 and 0.68, respectively, lower than those observed for each residential area individually. Furthermore, the verification outcomes of the Heibara district images using the TIr model (Table \u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003e) and the Tochihara I district images using the HIr model (Table \u003cspan class=\"InternalRef\"\u003e11\u003c/span\u003e) are provided below. Verification accuracy and overall F1-score for the Heibara district using the TIr model were 0.50 and 0.47, respectively. Similarly, verification results for the Tochihara I district using the HIr model yielded an accuracy and overall F1-score of 0.51. Both demonstrated lower accuracy compared to individual district verifications, suggesting that models tailored to each region were developed due to training solely on images from their respective areas. Moreover, the construction and evaluation of a joint model with a limited number of districts may have failed to capture common features among persimmon trees. Consequently, the concept of building an irrigation decision model adaptable to various fields by acquiring and training images from multiple fields warrants further investigation. Thus, future endeavors should delve into the feasibility of this approach, particularly concerning the number of fields involved.\u003c/p\u003e\n\u003cp\u003eTable 12 presents the proportion of misclassified images from each camera installed in the respective districts. Notably, for Flame 4 in the Tochihara I district and Flame 2 and Flame 4 in the Heibara district, which captured the entire tree, the proportions were notably high, at 35%, 43%, and 27%, respectively. As persimmon trees mature, their branches gradually bend under the weight of the fruit. Consequently, in Flame 2 of the Heibara district, initially, roughly half of the image was occupied by leaves. However, as time progressed, about half of the image was dominated by the ground. This resulted in a decrease in the proportion of leaves in the image, leading to an increase in misclassifications. Conversely, Flame 1 and Flame 3 in the Heibara district, which closely captured the leaves, exhibited lower proportions of misclassified images, at 13% and 17%, respectively, compared to images capturing the entire tree. Hence, it is preferable to capture images from locations where the angle of the leaves in the image remains consistent over time, facilitating maintenance. Furthermore, images capturing the entire tree had a lower proportion of misclassifications when leaves occupied a larger proportion of the image, suggesting that classification is easier when leaves dominate the image. Additionally, misclassified images occasionally featured unclear features due to backlighting or water droplets on the camera after rainfall. While these images were included in the training dataset, excluding them from training and evaluation is anticipated to enhance classification accuracy.\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e and \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e depict the decision criteria of the models following the methodology outlined by Zhou et al. (\u003cspan class=\"CitationRef\"\u003e2016\u003c/span\u003e) and Selvaraju et al. (\u003cspan class=\"CitationRef\"\u003e2017\u003c/span\u003e). The color bar in the figures represents normalized feature weights, with red indicating higher importance for classification and blue indicating lower contribution. Notably, in images capturing the entire tree in both districts, red heatmaps are predominantly concentrated on leaves closer to the camera, while blue heatmaps, indicating lower importance, are more prevalent for distant leaves. Furthermore, for images capturing leaves closely in the Heibara district, the heatmaps tend to be concentrated on the side of the leaves. This observation aligns with the common knowledge that leaves tend to wilt or curl when experiencing water stress, suggesting that the models may be leveraging this physiological characteristic to ascertain the necessity of irrigation.\u003c/p\u003e\n\u003cp\u003eIn summary, to enhance the effectiveness of the models, it is essential to install cameras closer to the leaves during image capture and utilize installation methods capable of capturing changes in leaf shape.\u003c/p\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eIn this study, by constructing models to estimate soil moisture from leaf images and to determine the necessity of irrigation, we were able to establish a correlation between leaf images and soil moisture levels, suggesting the potential for new irrigation indicators. While the accuracy of the soil moisture estimation model was likely affected by imbalances in the number of samples for each class, a highly accurate model was constructed for classes with a sufficient number of samples.\u003c/p\u003e \u003cp\u003ePrevious studies have reported that increasing irrigation improved fruit growth, but excessive soil moisture led to increased physiological fruit drop (Matsuda et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). Additionally, relationships have been found between irrigation levels and fruit size (Matsumura, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e1999\u003c/span\u003e), as well as between irregular fruit shapes and the proportion of unsuitable fruit for shipping (Kanety et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Therefore, precise irrigation management is necessary to improve quality and increase yield. Consequently, more detailed estimation of soil moisture is crucial.\u003c/p\u003e \u003cp\u003eIn the irrigation decision model, high-accuracy models were constructed for eachdistrict. However, the CIr model, blending images from two fields, exhibited low accuracy. Additionally, verification results from the Heibara district using the TIr model and the Tochihara I district using the HIr model displayed low accuracy, suggesting the development of specialized models for each field, possibly due to incomplete understanding of persimmon's general characteristics. Consequently, future endeavors should prioritize constructing versatile models that do not specialize in each field by augmenting the number of training samples and quantitatively evaluating the models' decision criteria.\u003c/p\u003e \u003cp\u003eThis model holds promise for facilitating remote irrigation decisions by integrating field-acquired images into edge devices for on-site processing. Moreover, when combined with remote irrigation systems currently in development, it can pave the way for automated irrigation, offering significant potential for labor-saving measures.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eA.O. wrote the main manuscript text and program coding, Data curation. A.Y. Data curation, M.K. Conceptualization, Methodology development, Y.M. Conceptualization, Methodology development, Research organization, Editing. All authors reviewed manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e \u003cp\u003eThis research received support from the development and improvement program of strategic smart agricultural technology grants provided by the Project of the Bio-oriented Technology Research Advancement Institution (BRAIN). The authors extend their sincere appreciation to Nara Prefecture for their support and facilitation of the research activities.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAoshima, W., Kawamura, S., Nakano, S., \u0026amp; Nakamura, H. (2018). Study on Variant Extraction of concrete structures using Image Recognition by Deep Learning. \u003cem\u003eThe Journal of the Japan Society of Civil Engineers E2 (Materials and Concrete Structures), 74(4)\u003c/em\u003e, 293-305. (In Japanese) https://www.jstage.jst.go.jp/article/jscejmcs/74/4/74_293/_pdf/-char/ja\u003c/li\u003e\n\u003cli\u003eBallester, C., Jim\u0026eacute;nez-Bello, M. A., Castel, J. R., \u0026amp; Intrigliolo, D. S. (2013). Usefulness of thermography for plant water stress detection in citrus and persimmon trees. \u003cem\u003eAgricultural and forest Meteorology, 168\u003c/em\u003e, 120-129. https://www.sciencedirect.com/science/article/pii/S0168192312002572\u003c/li\u003e\n\u003cli\u003eGojo-Yoshino Land Improvement District. Leaflet of Gojo-Yoshino Land Improvement District; Gojo-Yoshino Land Improvement District: Nara, Japan, 2002. (In Japanese)\u003c/li\u003e\n\u003cli\u003eGonzalez-Dugo, V., Zarco-Tejada, P., Nicol\u0026aacute;s, E., Nortes, P. A., Alarc\u0026oacute;n, J. J., Intrigliolo, D. S., \u0026amp; Fereres, E. J. P. A. (2013). Using high resolution UAV thermal imagery to assess the variability in the water status of five fruit tree species within a commercial orchard. \u003cem\u003ePrecision Agriculture, 14\u003c/em\u003e, 660-678. https://link.springer.com/article/10.1007/s11119-013-9322-9\u003c/li\u003e\n\u003cli\u003eIwata, Y., Nakagawa, F., Aida, N., Nawa, N., Miyamoto, T., Kameyama, K., \u0026amp; Syoubu, J. (2019). Effect of Drip Irrigation on the Weight of Fruits of Japanese Persimmon (Diospyros kaki Thunb.) at the Field in Sado Island, Japan. \u003cem\u003eTransactions of the Japanese Society of Irrigation, Drainage and Rural Engineering, 87(2)\u003c/em\u003e, I_227 - I_237. (In Japanese) https://www.jstage.jst.go.jp/article/jsidre/87/2/87_I_227/_article/-char/ja/\u003c/li\u003e\n\u003cli\u003eJapan Meteorological Agency (JMA). (2023). https://www.data.jma.go.jp/obd/stats/etrn/index.php (in Japanese).\u003c/li\u003e\n\u003cli\u003eKamoda, S. (1987): Environmental Control in Facility Cultivation of Fruit Trees. \u003cem\u003eAbstracts of Symposium of the Horticultural Society\u003c/em\u003e, 1-8. (In Japanese)\u003c/li\u003e\n\u003cli\u003eKanety. T., Naor A., Dicken. U., Lemcoff H., Cohen. S. (2014). Irrigation influences on growth, yield, and water use of persimmon trees. \u003cem\u003eIrrigation Science, 32\u003c/em\u003e, 1-13. (In Japanese)https://link.springer.com/article/10.1007/s00271-013-0408-y\u003c/li\u003e\n\u003cli\u003eTanaka, K., \u0026amp; Aoki, M. (1971). Effect of Irrigation and Nitrogen Application in Summer on the Fruiting of \u0026lsquo;Fuyu\u0026rsquo; Kaki. \u003cem\u003eResearch bulletin of the Aichi-ken Agricultural Research Center. Series B, 3\u003c/em\u003e, 9-18. (In Japanese)\u003c/li\u003e\n\u003cli\u003eLiu. W., Qin. C., Gao. K., Li. H., Qin. Z., Cao. Y., Si. Wen. (2019). Research on Medical Data Feature Extraction and Intelligent Recognition Technology Based on Convolutional Neural Network. \u003cem\u003eIEEE Access, 7\u003c/em\u003e, 150157 \u0026ndash; 150167. https://ieeexplore.ieee.org/document/8849989\u003c/li\u003e\n\u003cli\u003eMatsuda, M., Habu, T., Konishi, T., Nonaka, K., Kusumi, K., Kurosawa, T., \u0026amp; Kitajima, A. (2011). Effect of the drip irrigation system with plastic mulch on fruit quality in Japanese persimmon cv. \u0026lsquo;Taisyu\u0026rsquo;. \u003cem\u003eBulletin of the Experimental Farm, Kyoto University, 20\u003c/em\u003e, 29-32. (In Japanese) https://agriknowledge.affrc.go.jp/RN/2030830743\u003c/li\u003e\n\u003cli\u003eMatsuda, M., Matsumoto, D., Konishi, T., Nonaka, K., Kusumi, K., Kurosawa, T., \u0026amp; Kitajima, A. (2013). Effect of the drip irrigation system with plastic mulch on fruit quality in Japanese persimmon cv. \u0026lsquo;Taisyu\u0026rsquo;. Ⅱ.\u003cem\u003e Bulletin of the Experimental Farm, Kyoto University, 22\u003c/em\u003e, pp27-32. (In Japanese) https://agriknowledge.affrc.go.jp/RN/2030872756\u003c/li\u003e\n\u003cli\u003eMatsumura, H. (1999). Container Planting Cultivation of Japanese Persimmon (Diospyros kaki Thunb.) (Pat3) -The Improvement of the Fruit Quality by Fertilization and irrigation. \u003cem\u003eJournal of Society of High Technology in Agricultur\u003c/em\u003ee\u003cem\u003e, 11(4)\u003c/em\u003e, 259-266. (In Japanese) https://www.jstage.jst.go.jp/article/jshita1991/11/4/11_4_259/_article/-char/ja/\u003c/li\u003e\n\u003cli\u003eMuramatsu, N., \u0026amp; Hiraoka, K. (2006). Estimation of Water Content in the Leaves of Fruit Trees Using Infra-red Images. \u003cem\u003eHorticultural Research, 5(4)\u003c/em\u003e, 397-402. (In Japanese) https://www.jstage.jst.go.jp/article/hrj/5/4/5_4_397/_article/-char/ja/\u003c/li\u003e\n\u003cli\u003eOkatani, T. (2022). Deep Learning, \u003cem\u003eRevised 2nd Edition. Kodansha\u003c/em\u003e. (In Japanese)\u003c/li\u003e\n\u003cli\u003eO\u0026apos;Shea, K., \u0026amp; Nash, R. (2015). An introduction to convolutional neural networks. \u003cem\u003earXiv preprint.\u003c/em\u003e (In Japanese) https://arxiv.org/abs/1511.08458\u003c/li\u003e\n\u003cli\u003eSarma, K. K., Das, K. K., Mishra, V., Bhuiya, S., \u0026amp; Kaplun, D. (2022). Learning Aided System for Agriculture Monitoring Designed Using Image Processing and IoT-CNN. \u003cem\u003eIEEE Access, 10\u003c/em\u003e, 41525-41536. https://ieeexplore.ieee.org/document/9756536\u003c/li\u003e\n\u003cli\u003eSelvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., \u0026amp; Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. \u003cem\u003eIn Proceedings of the IEEE international conference on computer vision,\u003c/em\u003e 618-626. https://arxiv.org/abs/1610.02391\u003c/li\u003e\n\u003cli\u003eShermila, P. J., Victor, A., Manoj, S. O., \u0026amp; Devi, E. A. (2024). Automatic detection and classification of disease in citrus fruit and leaves using a customized CNN based model. \u003cem\u003eBolet\u0026iacute;n Latinoamericano y del Caribe de Plantas Medicinales y Arom\u0026aacute;ticas, 23(2)\u003c/em\u003e, 180-198. https://www.blacpma.ms-editions.cl/index.php/blacpma/article/view/409/408\u003c/li\u003e\n\u003cli\u003eZhou, B., Khosla, A., Lapedriza, A., Oliva, A., \u0026amp; Torralba, A. (2016). Learning deep features for discriminative localization. \u003cem\u003eIn Proceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e, 2921-2929. https://arxiv.org/abs/1512.04150\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 12 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"environmental-monitoring-and-assessment","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"emas","sideBox":"Learn more about [Environmental Monitoring and Assessment](http://link.springer.com/journal/10661)","snPcode":"10661","submissionUrl":"https://submission.nature.com/new-submission/10661/3","title":"Environmental Monitoring and Assessment","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Persimmon irrigation, Convolutional Neural Networks, Decision-making, Remote sensing, Automated irrigation","lastPublishedDoi":"10.21203/rs.3.rs-4404121/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4404121/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study aimed to develop and evaluate a model for persimmon irrigation decision-making using Convolutional Neural Networks (CNNs) based on leaf image data. Data comprising leaf images collected in the field and corresponding soil moisture measurements were gathered from the Gojo Yoshino region, recognized as the primary persimmon-producing area in Nara Prefecture, Japan's second-largest persimmon-producing prefecture. The findings demonstrate that the constructed CNN model can successfully identify water stress levels in persimmon trees from leaf image data. However, there are limitations to the model's performance and scopes for improving accuracy. The model's capability enables remote irrigation decision-making by integrating field-acquired leaf images into edge devices for on-site processing. When integrated with ongoing developments in remote irrigation systems, this technology has the potential to automate irrigation practices, thereby offering substantial labor-saving benefits.\u003c/p\u003e","manuscriptTitle":"Utilizing Convolutional Neural Networks (CNN) for Persimmon Irrigation Decision-Making: A Case Study in the Gojo Yoshino Region, Nara Prefecture, Japan","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-06-03 13:55:29","doi":"10.21203/rs.3.rs-4404121/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-06-06T23:15:15+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-06-06T23:01:26+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-06-06T09:59:14+00:00","index":"","fulltext":""},{"type":"submitted","content":"Environmental Monitoring and Assessment","date":"2024-05-11T07:32:23+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"environmental-monitoring-and-assessment","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"emas","sideBox":"Learn more about [Environmental Monitoring and Assessment](http://link.springer.com/journal/10661)","snPcode":"10661","submissionUrl":"https://submission.nature.com/new-submission/10661/3","title":"Environmental Monitoring and Assessment","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"a9bb2844-a68f-4ead-a3ac-0cfadfa883e0","owner":[],"postedDate":"June 3rd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-01-20T15:59:42+00:00","versionOfRecord":{"articleIdentity":"rs-4404121","link":"https://doi.org/10.1007/s10661-024-13602-1","journal":{"identity":"environmental-monitoring-and-assessment","isVorOnly":false,"title":"Environmental Monitoring and Assessment"},"publishedOn":"2025-01-14 15:57:01","publishedOnDateReadable":"January 14th, 2025"},"versionCreatedAt":"2024-06-03 13:55:29","video":"","vorDoi":"10.1007/s10661-024-13602-1","vorDoiUrl":"https://doi.org/10.1007/s10661-024-13602-1","workflowStages":[]},"version":"v1","identity":"rs-4404121","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4404121","identity":"rs-4404121","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}