An agricultural digital twin for mandarins demonstrates the potential for individualized agriculture

The concept of a digital twin (DT), a digital representation mirroring a real-world object, has gained traction across various sectors, including agriculture. Leveraging big data and systematic data management, DTs offer predictive capabilities. This study aims to build an agricultural DT for mandarins, a model crop widely cultivated in Jeju Island, South Korea. The perennial nature of mandarin trees and their relatively large spacing facilitate the collection of individual-specific, longitudinal data crucial for spatiotemporal analysis. Data aggregation is achieved using an Open API to access publicly available datasets from multiple sources, including the Rural Development Administration (RDA) for soil data, the Jeju Free International City Development Center (JDC) for fruit quality and weather data, and the Ministry of the Interior and Safety (MOIS) for geocoding. This comprehensive approach allows for data visualization and analysis at various scales – regional, inter-orchard, and intra-orchard. The study aims to demonstrate how this DT supports data-driven decisions, transitioning from precision agriculture towards individualized agriculture, where each tree receives customized treatment.

Existing literature highlights the use of DTs in various fields such as aerospace, automotive, manufacturing, and healthcare. In agriculture, previous studies have explored the application of IoT-based platforms for data collection (SmartFarmNet), UAV-based intrusion detection, orchard-scale DTs using 3D LiDAR for tree health monitoring, WebGIS frameworks for big data analytics, and deep learning algorithms for orchard tree segmentation. These studies demonstrate the feasibility of DT applications but often focus on specific aspects or scales. This research builds upon these foundations, aiming for a more comprehensive and integrated DT for open-field fruit crop management, addressing the need for individualized agricultural practices at the tree level.

Data were collected using publicly accessible Open APIs from various sources on the data portal (https://www.data.go.kr). The data included soil chemical properties (from the RDA), fruit quality, weather information, and agricultural practices (from the JDC), and geocoding information (from the MOIS). Data parsing and analysis were conducted using R. Geocoding was performed using the Kakao Developers server. Regional-scale analysis employed kernel density estimation (KDE) to visualize soil components across Jeju Island. Inter-orchard analysis compared two orchards (Hab and lab) with the same cultivar, examining soil properties, agricultural practices, and fruit quality. Intra-orchard analysis focused on the lab orchard, analyzing fruit quality variations within the orchard using hierarchical clustering and mixed-effects models. An automatic machine learning (AutoML) algorithm, implemented using the 'h2o' package in R, was used to predict fruit size and sugar content based on various orchard-level variables. An interactive applet, created using R Shiny, was developed to visualize the DT's outputs.

Regional-scale analysis revealed variations in soil components across Jeju Island, with higher levels of certain nutrients in the western region compared to the east. Inter-orchard analysis highlighted significant differences in fruit quality between the Hab and lab orchards, despite the same cultivar, attributed to differences in soil properties and agricultural practices. Intra-orchard analysis of the lab orchard showed considerable variability in fruit quality among trees, indicating the potential for individualized treatment. The AutoML algorithm showed higher predictive accuracy for fruit size than for sugar content at the orchard level, suggesting the need for intra-orchard analysis to improve sugar content prediction. The mixed-effects models indicated that while orchard-level variables explained some variance in fruit quality, a tree-level model significantly improved the prediction of sugar content (R² increased from 0.19 to 0.66). This highlights the significant contribution of intra-orchard variability to sugar content.

The findings demonstrate the potential of agricultural DTs for individualized agriculture. The ability to predict fruit quality at the tree level offers opportunities for precise interventions, leading to customized agricultural practices that maximize yield and quality. This approach moves beyond the limitations of precision agriculture, which primarily focuses on orchard-level management. The significant increase in R² when moving from orchard-level to tree-level analysis underscores the importance of considering the intra-orchard variability for optimal fruit quality management. The interactive applet provides a user-friendly platform for visualizing and utilizing the data for decision-making, highlighting the practicality of this approach for various stakeholders.

This study successfully demonstrates the development and application of an agricultural DT for mandarin orchards. The results highlight the potential of individualized agriculture, where customized management practices are applied to individual trees to maximize fruit quality. Future work should focus on expanding data collection, refining predictive models, and developing cost-effective strategies for implementing DT-guided management in various fruit crops and even extending to other crop types.

The study was limited by the availability of data. While the use of publicly available datasets through Open APIs is a strength, the data were not collected specifically for this research, leading to potential biases and limitations in the analysis. The currently available applet does not automatically suggest agricultural practices nor assess their effects on fruit quality. This requires further development and real-world feedback from stakeholders. The study focused on mandarins; the generalizability to other crops needs further investigation.