Conference Agenda

Groundwater-depth forecasts are increasingly required to support irrigation management and nitrate-risk assessment in Mediterranean agro-hydro-systems, where shallow alluvial aquifers exhibit strong seasonal variability and rapid responses to pumping and recharge. Within the scope of the Clepsydra Project, this study presents an operational, well-specific forecasting workflow applied to the Tagus Nitrate Vulnerable Zone in central Portugal. Monthly groundwater depth is modelled independently at each monitoring well using an autoregressive–exogenous (ARX) model trained with gradient boosting (XGBoost). The workflow explicitly addresses data gaps and heterogeneous record lengths and is evaluated using leakage-safe rolling-origin validation for forecast lead times ranging from 1 to 12 months. Forecast performance is assessed using conservative error metrics computed exclusively on observed (non-imputed) values. Results demonstrate stable and competitive accuracy up to a 12-month horizon, with a gradual and near-linear degradation in performance as lead time increases. Post hoc analyses indicate that groundwater persistence and seasonal components are the dominant predictors, while hydro climatic drivers contribute in a site-dependent manner. The 12-month-ahead forecasts remain physically plausible and consistent with historical dynamics, supporting their integration into groundwater monitoring and decision-support systems for irrigation planning and nitrate regulation

Cultivating crops under artificial lighting in smart greenhouses requires precise environmental control. Since light intensity fluctuations directly impact crop growth, reliable data collection is paramount. However, in real operational environments, data quality often degrades due to factors like sensor failures or communication instability. These missing values and outliers significantly hinder the stability of lighting control and the performance of AI-based predictive models.

To address these challenges, this study designed an AI-driven automated data quality detection and correction system. By adopting a microservice architecture, we maximized scalability and management efficiency. The application of Isolation Forest and AutoEncoder algorithms confirmed the potential for superior detection accuracy compared to conventional methods. This system minimizes human intervention through automated monitoring and is expected to contribute to enhanced operational and energy efficiency. By proposing an intelligent data management framework, this research establishes a technical foundation for improving the reliability of the smart farm ecosystem. Future work will focus on edge computing, Transformer-based model expansion, and root-cause analysis for anomalies.

With the advancement of IoT technology and the expansion of smart farming, the utilization of vast environmental data has become increasingly critical for integrated intelligent diagnostic systems. However, conventional CNN-based models are primarily limited to visual information, failing to organically incorporate complex environmental contexts and lacking the explainability required for agricultural decision-making.

To address these limitations, this study develops an ontology-based multimodal Retrieval-Augmented Generation (RAG) agent. The system architecture consists of three core modules: 1) a Vision-Language Model (VLM) that translates crop images into specialized agricultural terminology to enhance semantic search precision; 2) a lightweight knowledge base that optimizes ontology schemas into vector database metadata for computational efficiency; and 3) an environmental context inference engine that filters out logically inconsistent diagnostic candidates by cross-referencing real-time temperature and humidity data with disease occurrence conditions.

Experimental results demonstrate that the proposed system achieves a significant improvement in diagnostic accuracy compared to standard image-similarity searches. By operationalizing theoretical ontologies into a practical architecture, this research provides a robust reference model for Explainable AI (XAI) in agriculture. Furthermore, the system is expected to minimize economic losses caused by misdiagnosis and contribute to enhanced agricultural productivity through data-driven precision pest management.

Soil compaction is recognized as a major driver of structural soil degradation, reducing porosity, restricting root development, and ultimately affecting crop yield and agricultural profitability. In this context, this study aimed to evaluate the feasibility of relating UAV-derived vegetation parameters to soil penetration resistance and digitally mapped soil properties for spatial soil compaction assessment in a vineyard.

Vegetation indices (VIs), canopy area (CA), and canopy volume derived from dense point cloud reconstruction were extracted from multispectral UAV imagery. Soil penetration resistance (SPR) was measured down to 0.60 m using a portable digital penetrometer. Soil texture and organic matter (OM) were mapped using the Veris iScan system equipped with a dual-wavelength infrared soil optical sensor. Spatial and multivariate statistical analyses were performed to investigate relationships between canopy parameters and depth-dependent SPR profiles.

Significant differences in canopy area and canopy volume were detected across management zones characterized by contrasting penetration resistance and OM-related soil texture. Correlation analysis showed that selected VIs were significantly associated with compaction-related parameters, supporting the indirect estimation of soil resistance from aerial data.

The integration of UAV multispectral sensing and proximal soil measurements provides a scalable framework for precision soil management within digital farming systems.

Precision oliviculture aims to improve yield and product quality while reducing production costs and environmental impacts. In olive orchards, balanced nitrogen management based on the actual spatial variability of soil and plant conditions is essential to maintain a proper vegetative–productive equilibrium. This study aimed to characterise soil, canopy, and yield variability within an intensive olive orchard using an integrated multi-sensor approach in order to generate a site-specific nitrogen fertilisation prescription map. The experiment was conducted in an intensive olive orchard located in a Mediterranean environment. Nitrogen requirements were estimated by considering both soil nitrogen availability and crop nitrogen removal associated with fruit yield and pruning biomass. UAV-based multispectral imagery was acquired at 70 m above ground level to evaluate canopy spectral variability. Yield spatial distribution was obtained from yield-bin data collected during mechanical harvesting, while soil variability was assessed using a Veris iScan sensor combined with ground-truth soil sampling. All datasets were integrated within a GIS environment to produce a variable-rate nitrogen fertilisation prescription map. The proposed approach enabled the definition of site-specific nitrogen management zones, achieving a 27% reduction in fertiliser use compared with conventional uniform-rate application, while maintaining production levels and improving environmental sustainability.

Modern agriculture is increasingly challenged by growing food demand, water scarcity, and the impacts of climate change, making efficient water resource management a key priority for sustainable production systems. In this context, smart farming technologies and digital monitoring systems represent valuable tools for improving irrigation management and reducing environmental impact.

This research focuses on the design, calibration, and validation of a low-cost monitoring system for soil moisture and temperature, using environmental sensors (DHT22 and SHT10) and soil sensors (SEN0308 and DS18B20) connected to a microcontroller for continuous acquisition of environmental and soil parameters. Soil moisture was estimated from capacitive sensor readings converted into volumetric water content (VWC) through a gravimetric calibration procedure. Statistical analyses were conducted to evaluate sensor performance, including normality tests, and correlation analyses.

The calibration process resulted in a soil-specific second-order polynomial equation implemented directly in the Arduino code, significantly improving the accuracy of soil moisture estimation.

The system was also connected to a FastAPI-based application programming interface, enabling structured data access and preparing the system for remote monitoring and decision-support applications. Overall, the results demonstrate that properly calibrated low-cost sensors can provide reliable and scalable solutions for real-time soil water monitoring and sustainable irrigation management.

In Costa Rica, many Water User Associations (Sociedades de Usuarios de Agua, SUAs) emerged as a response to historically informal and, in some cases, illegal water withdrawals. Their formalization represented a major step toward transitioning from unregulated, fragmented practices to organized, community-based irrigation management. However, today’s pressures, climate variability, increasing competition for water, and the need for higher efficiency demand a deeper transformation that goes beyond organizational legalization.

This poster presents a digital transformation roadmap grounded in a digital governance model to strengthen not only operational efficiency but also transparency, accountability, and evidence-based decision-making in community irrigation schemes. We argue that digital transformation is not simply the rollout of technology; it is about keeping users at the center, reinforcing institutional coordination, and ensuring that data flows, rules, and decision processes are clearly defined and collectively legitimized.

A key message is that managing the technological architecture for irrigation modernization must be paired with capacity building, training, and social appropriation so that digital tools become sustainable and are effectively used by farmers and local leaders. We propose that this governance-oriented approach offers an inclusive, competitive, resilient, and replicable framework for similar community-managed irrigation systems in Costa Rica and across Latin America.

The “Impact of Agricultural Digitalization for the CAP Network” (DIGIPAC program), funded by the Italian Ministry of Agriculture, Food Sovereignty and Forestry (MASAF) through the 2023-2027 CAP Network, is a strategic initiative aimed at fostering digital agriculture in Italy. It focuses on promoting technological innovation and the integration of advanced digital tools within the national farming sector. By facilitating knowledge transfer and the adoption of smart farming practices, DIGIPAC supports the sustainable growth and competitiveness of Italian agribusiness. This program provides a comprehensive overview of the technological solutions in different production sectors, the current adoption of AI, machine learning, and big data analytics, including autonomous field operations via robotic guidance and real-time monitoring through IoT sensor networks and edge computing, to improve crop quality while minimizing the input use and environmental impacts. Smart farming strategies leverage cloud systems and AI/DSS to enable predictive analysis and resource optimization. The digitalization of supply chains through intelligent platforms ensures data governance and enhanced traceability. The central objective of this research is to address existing knowledge gaps by establishing a systematic framework for agricultural data standards and statistical collection. By mapping the state of technology adoption, the initiative provides evidence-based insights for policymakers and stakeholders.

Spatio-temporal fusion techniques, integrating UAV and Sattelite, enable accurate parcel-level crop monitoring with high spatial, temporal resolution particularly under rapid phenological changes. This study was conducted over a soybean field located in Anseong, Gyeonggi-do, Republic of Korea. We compared modified STARFM, Fit-FC, FSDAF, UBDF, and RASTFM algorithms, and further assessed the impact of CACAO post-processing on prediction accuracy using UAV observations as reference data. Among the evaluated models, the modified STARFM algorithm achieved the highest overall accuracy, and the application of CACAO smoothing significantly improved predictive performance across most spectral bands and vegetation indices. These findings are critical for reliable intra-field spatial-temporal crop monitoring and can support precision agriculture and data-driven farm management practices.

One of the main objectives of crop management is effective weed control. This work presents an intelligent weed mapping approach for patch spraying applications, based on unmanned aerial vehicle (UAV) imagery and deep learning techniques. Several models were developed to automatically detect and identify four common weed species found in tomato cultivation fields (Amaranthus retroflexus, Chenopodium album, Cyperus esculentus and Portulaca oleracea). The models were trained on a dataset of 1064 RGB images captured at various heights with a general-purpose UAV. The resulting models, assessed on a separate test set, performed satisfactorily, achieving sufficient detection accuracy across all weed classes (Precision 0.776, Recall 0.777, mAP₅₀ 0.828).

The developed weed identification system was applied to UAV imagery captured during a new growing season, with a goal to generate up-to-date weed distribution maps. These maps provide spatially explicit information on weed locations and densities, forming the basis for site-specific herbicide application through patch spraying.

The proposed method, by reducing unnecessary herbicide use, supports a more sustainable weed management, targeting weeds with high spatial precision. It offers a scalable approach to weed control, guided by real-time data, incorporating artificial intelligence methodologies, use of non-specialized UAV equipment, and relatively simple geographic information systems techniques.

Effective pest and disease management is a critical aspect of precision agriculture. Unmanned aerial spraying systems (UASS) have been increasingly adopted in crop protection, but challenges remain, including the high cost of precision spraying platforms and uneven spray distribution in conventional systems. This study presents a low-cost and modular autonomous spraying control system composed of a ground base station and an onboard control assistant. A spray uniformity control method based on a regression forest model is implemented to maintain the coefficient of variation (CV) of spray distribution below 30%. Real-time environmental data are also collected to support adaptive spraying adjustments. Environmental data and positioning correction signals are transmitted from the ground base station to the onboard module via LoRa communication, enabling accurate positioning and dynamic control during spraying operations. Indoor simulation experiments show that the system maintains spray distribution within the standardized CV requirement (<30%). Field experiments using a simulated prescription map further demonstrate that the system can successfully perform targeted spraying across designated zones, indicating its potential to improve UASS spraying precision and efficiency.

Small UAVs relying solely on optical-flow sensors struggle to achieve the precision hovering required in confined indoor agricultural environments without GPS. This study presents an autonomous flight system integrating AprilTag-based visual servoing with companion-grade onboard hardware and evaluates its stability in a GPS-denied greenhouse setting. The proposed system corrects the error between the marker center and the image frame through a PID-based outer control loop, directly commanding flight control signals to align the UAV within predefined tolerance ranges before proceeding to subsequent operations. The platform combines a Pixhawk flight controller running ArduPilot firmware with a Raspberry Pi 5, stabilized by low-cost optical-flow and LiDAR sensors. To validate precision improvement over conventional hovering, the system is compared against pure LOITER mode, which relies exclusively on optical-flow without external correction. Experimental results show an alignment success rate of 39.6% and a mean alignment time of 1.73 s, with RMS vibration of 8.62 and lateral oscillation of 28.9 px, representing significant improvement over LOITER-only operation (26.0%, 2.71 s) and a 28.3% reduction in lateral error. Further validation under varied greenhouse lighting is expected to broaden applicability in controlled-environment agriculture.

Unmanned aerial vehicles are increasingly used in precision agriculture, yet the operational chain from raw imagery to reusable artificial-intelligence services often remains fragmented across separate tools for image processing, dataset preparation and model deployment. This work presents the SPADE Toolbox, an open modular workflow designed to connect these stages into an interoperable end-to-end pipeline for UAV-enabled agricultural intelligence. The toolbox integrates complementary components for cloud-based processing of RGB and multispectral drone imagery into geospatial products, generation and annotation of machine-learning-ready agricultural datasets, and containerized deployment of AI models as reusable API-accessible services. The workflow supports the transformation of raw UAV data into orthomosaics, digital surface models, point clouds, vegetation-index products and structured training resources, while enabling human-supervised semi-automatic annotation and scalable execution of image-analysis models. The resulting architecture demonstrates how heterogeneous digital services can be organized into a coherent toolbox that supports reproducibility, interoperability and extensibility across the full data-to-service chain. By reducing the gap between UAV data acquisition and operational AI applications, the SPADE Toolbox provides a practical framework for developing reusable digital services for precision agriculture and related agri-environmental domains

Cultivating Medjool date palms requires careful management. Fruit thinning removes excess developing fruit to ensure high yields of high-quality fruit. Growers make complex decisions in dynamic conditions, often relying on fragmented/inconsistently recorded information. Model-based decision-support systems can improve farm management by integrating and analyzing data from multiple sources.

This study aimed to predict fruit yield and quality as influenced by thinning decisions and climate variables. Meteorological data, annual yield records, and thinning intensity, estimated through fruitlet counts, were compiled from 2013 to 2023. Individual fruits were categorized by skin separation and weight class. Four machine-learning models (Ridge, ElasticNet, Random Forest, and XGBoost) were evaluated to predict yield per tree across prediction horizons and feature sets. Dirichlet regression models were developed to predict compositional outcomes of skin separation and fruit weight.

XGBoost using coverage-stage counts and climate variables achieved the best performance (R²=0.70±0.12; RMSE=21.3±2.0 kg/tree). Skin separation and fruit weight models improved prediction accuracy by up to 20.6% and 11.8%, respectively, over baseline. The models developed in this study highlight the dominant influence of thinning decisions on yield outcomes, while meteorological variables contributed limited additional predictive value. In contrast, a clear relationship was identified between climatic conditions and skin separation.

This paper presents the concept and development of an integrated approach for assessing potato crop productivity and quality by combining satellite observations, climate data, and field measurements. The aim of the study is to improve the early identification of plant stress and the risk of yield reduction through the integration of multiple data sources. The system utilizes Sentinel-2 satellite imagery to derive vegetation and moisture indices that provide information on crop condition, nitrogen status, and water stress. In addition, climate indicators derived from ERA5 data—including air temperature, soil moisture, solar radiation, and heat stress parameters—are used to characterize environmental conditions affecting plant growth. The collected data are processed within a cloud-based infrastructure and made available through web interfaces and mobile applications designed for both users and administrators. Integrated analysis enables the estimation of potato crop productivity at the field level before harvest and supports the identification of spatial variability, stress zones, and potential risks affecting crop quality. Initial results suggest that combining satellite-derived vegetation indices with climate data may provide a valuable tool for supporting potato crop monitoring and improving crop management in precision agriculture.

Variability in crop growth persists even under identical cultivation conditions, highlighting limitations of conventional smart farming systems based on uniform, average-based management. Such approaches often fail to account for plant-level heterogeneity caused by microenvironmental and physiological differences, leading to inefficient resource use and reduced productivity. Although customized management is essential for next-generation smart farming, practical and non-invasive methods for monitoring soil conditions at the individual plant scale remain limited. This study proposes an AI-based framework that integrates soil sensor measurements and image-based analysis to enable customized soil moisture management. The target plant of this study is transplanted wild-simulated ginseng. The RGB images of the soil surface were collected with in-situ soil moisture sensing and monitoring data at multiple depths to characterize vertical moisture dynamics. Machine learning models were systematically evaluated for moisture estimation. DenseNet121 achieved the highest accuracy for surface soil moisture prediction, while random forest regression showed superior performance for subsurface layers by capturing nonlinear moisture behavior. The results demonstrate that surface RGB imagery can serve as a reliable, non-invasive proxy for soil moisture estimation when combined with data-driven models. The proposed system supports plant-specific decision-making and provides a scalable approach for precision moisture management in smart farming systems.

Accurate evaluation of tomato color is essential for determining fruit maturity, quality, and market value in smart agriculture and post-harvest management. Objective identification of color type, combined with quantitative estimation of color area ratio and HSL values, provides a reliable basis for grading, supports automated harvesting, and enables data-driven supply chain management. This study proposes an automated tomato color analysis system consisting of two main modules: (1) color type classification and (2) color area ratio estimation. Tomato fruits are first segmented from raw images using a Mask R-CNN model to eliminate background interference. The segmented images then undergo adaptive RGB (aRGB) preprocessing, which improves color consistency by mapping each pixel to the nearest reference in the QpCard color table using Euclidean distance. The normalized images are fed into a VGG16-based deep learning model with Spatial Pyramid Pooling and softmax classification to distinguish single-color and bi-color tomatoes. Subsequently, K-means clustering identifies dominant color regions and computes their corresponding area ratios. The system was evaluated on mature and immature tomatoes, achieving classification accuracies of 94% and 88%, respectively. These results demonstrate that the proposed framework provides an efficient and reliable solution for objective tomato quality assessment and intelligent grading applications.

Accurate strawberry weight estimation is essential for commercial production but is still largely performed manually, resulting in high labor costs and inconsistent assessments. Because fruit weight is strongly correlated with geometric traits such as width and height, this study developed a non-destructive RGB–D vision framework for automated weight prediction. Two deep learning approaches were evaluated for geometric feature extraction: instance segmentation using Mask R-CNN and keypoint detection using YOLOv8-Pose. Fruit dimensions were derived from segmented masks and detected keypoints and converted to real-world measurements using depth data. These geometric features were used to predict fruit weight through linear regression. Mask R-CNN achieved mean absolute error (MAE) values of 2.99 mm for width and 3.83 mm for height. Compared with YOLOv8-Pose, this represented a 2.5% reduction in height estimation error, although width estimation error increased slightly by 1.4%. Consequently, weight prediction improved, achieving R² = 0.86 and MAE = 1.48 g, compared with R² = 0.79 and MAE = 1.77 g for the keypoint-based method. Although estimation errors increased with fruit size, deviations remained within acceptable commercial grading ranges. Overall, pixel-wise boundary extraction provided a more reliable representation of fruit geometry than keypoint localization for robust automated strawberry weight estimation.

Unmanned Aerial Vehicle (UAV)-based image analysis enables precise identification of crop growth status using high-resolution spatial information. Specifically, Hyperspectral Imaging (HSI) facilitates the quantitative analysis of physiological and biochemical. This study aims to identify optimal vegetation indices (VIs) for classifying the growth stages of open-field garlic (Allium sativum L.) in Korea using UAV-based hyperspectral imagery.

Hyperspectral data were acquired at three key intervals: bulb expansion, bolting, and late bulb expansion. Geometric and radiometric calibrations ensured temporal data consistency, and pure-pixel garlic canopy spectra were extracted by masking soil background. Subsequently, various VIs were calculated to establish spectral indicators for growth stage discrimination.

Several machine learning models were implemented and compared for classification with performance evaluated via 5-fold cross-validation using accuracy, F1-score, and confusion matrices. Furthermore, feature importance analysis identified the most effective VIs. The results demonstrated that specific VIs exhibited high explanatory power for distinguishing the three growth stages, and the models utilizing these key indices achieved superior classification performance.

This study demonstrates the feasibility of garlic growth stage classification using UAV-based HSI and proposes a methodology for optimal vegetation index selection. These findings provide a baseline for garlic growth monitoring systems and future research into yield prediction and analysis.

In greenhouse strawberry production, harvesting and sorting require substantial labor, and post-harvest losses are often high. Non-destructive quality evaluation during the growth stage could reduce labor requirements by enabling pre-harvest sorting. Hyperspectral imaging is a promising technique for this purpose, and snapshot-type hyperspectral cameras are particularly suitable because they acquire spectral data without scanning. This study aimed to develop an algorithm for estimating soluble solids content (SSC) of strawberries using a snapshot-type hyperspectral imaging system under laboratory conditions. Strawberry fruits were measured under halogen illumination, and SSC reference values were obtained as Brix values using a refractometer. Image processing algorithms were developed to extract regions of interest corresponding to fruit surfaces, excluding pixels irrelevant to SSC estimation. These included sepal extraction using standard normal variate processing combined with thresholding, and achene extraction based on principal component analysis and image processing. Partial least squares regression models were constructed using the mean spectra of the extracted fruit surface regions, and spectral preprocessing conditions were optimized. The developed model achieved moderate predictive performance for SSC estimation. In addition, SSC distributions on the fruit surface were visualized as heatmaps. The proposed approach is expected to contribute to non-destructive quality evaluation of strawberries during growth.

Spectroscopy offers a non-destructive strategy for assessing lettuce water stress and supports robust stress-level classification across multiple experimental days. In this study, we developed a spectral machine-learning workflow using full-spectrum visible and near-infrared (VIS-NIR) data collected in the 500–900 nm range from lettuce plants subjected to progressive water deprivation. Spectra were preprocessed using Standard Normal Variate (SNV) normalization to remove multiplicative scatter effects and account for path-length variations. To avoid data leakage, spectra were split at the plant level (no plant appeared in both the training and test sets) while preserving class balance. Leave-one-out cross-validation (LOOCV) was applied across the entire dataset comprising spectra from 5 experimental days. Results showed strong discrimination performance. Partial Least Squares Discriminant Analysis (PLS-DA) achieved 96.3% accuracy under LOOCV, substantially outperforming Logistic Regression (81.5%), Random Forest (85.2%), and SVM-RBF (35.2%). PLS-DA's superior performance stems from its ability to extract latent spectral–stress relationships while simultaneously managing the high-dimensional collinearity inherent in full-spectrum data. These findings establish VIS-NIR spectroscopy combined with PLS-DA as an effective, non-destructive approach for real-time monitoring of lettuce stress. Future work should validate this framework across additional lettuce cultivars, alternative stress types, and greenhouse field conditions.

As the demand for agricultural automation increases, robotic harvesting systems capable of accurately perceiving crop conditions have become increasingly important. Reliable maturity classification is essential for automated strawberry harvesting. This study proposes a machine-learning-based strawberry maturity classification method using color-space features and analyzes the effect of different labeling strategies on classification performance.

A total of 2,800 images (8,990 strawberry fruits) were selected from an existing dataset, and maturity was categorized into three stages: unripe, semi-ripe, and ripe. Color features were extracted from RGB, HSV, and Lab color spaces, with outliers removed via box-plot analysis. Classification performance was compared between segmentation-based and bounding box-based labeling frameworks.

The a component of the Lab color space achieved a classification accuracy exceeding 0.85. With mixed color features, the SVM-based GLaS model achieved an accuracy of 0.908. Segmentation-based labeling, using the center axis defined by the two farthest points of the fruit, achieved an accuracy of 0.912 and a Kappa coefficient of 0.844, outperforming the bounding box-based approach (accuracy: 0.866, Kappa: 0.766). These results confirm that segmentation-based labeling is more effective for strawberry maturity classification and offers practical guidance for visual perception in autonomous harvesting robots.

Fresh-market tomato production is sensitive to in-field losses due to its continuous harvesting. Accurately quantifying yield losses is critical for supporting farmer decision-making. The objective of this study was to apply a convolutional neural network (YOLO) to high-resolution imagery to detect fruit losses during tomato harvesting. Images were acquired using a UAV DJI Matrice 300 equipped with a Sony ILCE A7R IV camera in a commercial tomato field in Brazil. The flight was conducted at a height of 30 m, producing a ground sample distance of 0.48 cm per pixel. A dataset of 120 images was divided into training, validation, and test subsets, and each tomato fruit on the ground was annotated and classified as unripe, semi-ripe, or ripe. After benchmarking different models, YOLOv8n was selected and applied to the test dataset, achieving AP values of 0.798, 0.725, and 0.582 for unripe, semi-ripe, and ripe fruits, respectively (mAP@0.50 of 0.702). The study also compared manual and automatic counting methods, obtaining an R² of 0.91 and a MAPE of 19.63%. Therefore, the proposed method effectively detected tomato fruits by class and can be used as a digital tool to monitor harvesting losses.

Diseases affecting Solanaceae crops such as early blight (Alternaria solani) and late blight (Phytophthora infestans), can cause major economic losses, especially when infections occur at advanced stages. Early, accurate, and non-destructive detection is therefore critical for precision agriculture, improved disease management, and reducing fungicide use. This study proposes a multispectral imaging approach combined with artificial intelligence to detect early and late blight infections in greenhouse-grown tomato leaves. Multispectral data were collected from healthy and artificially inoculated plants under controlled conditions, focusing on key spectral bands. A modified dual-head SegFormer architecture was used to perform semantic segmentation of healthy, early blight–infected, and late blight–infected leaf regions. Each dataset capture included five hyperspectral wavelength images (460, 540, 640, 780, and 880 nm) along with one RGB image. From these bands, five vegetation indices—NDVI, CVI, GNDVI, NPCI, and PSRI—were calculated to highlight disease-related spectral characteristics and enhance model performance. Experimental results showed that the dual-head SegFormer model achieved a mean Average Precision (mAP50) of 85% across all classes, demonstrating reliable detection and segmentation under varying disease conditions. The proposed approach shows strong potential for integration into automated greenhouse monitoring systems and robotic platforms for real-time crop health assessment.

Seed is one of the most critical factors in paddy production. Traditional techniques such as tetrazolium staining, conductivity tests, accelerated ageing tests, and germination tests for determining paddy seed viability are labour-intensive. To overcome these problems, this study proposes a deep neural network that works by fusing RGB images and corresponding image features to predict seed viability. Firstly, individual seed images were captured using a CCD camera. The seeds were then germinated according to the ISTA guidelines. A total of 7 colors, 9 morphological, and 4 textural features of the seed images were extracted. The deep neural network has two branches: one for image classification, containing 2 CNN layers, 2 MaxPooling layers, and 1 dense layer; and another for numerical classification, containing 2 dense layers with ReLU activation. The two branches are concatenated using a fully connected layer and Dropout for regularization. The proposed deep neural network was evaluated against classical machine learning algorithms, including Linear Discriminant Analysis, Support Vector Machine, K-nearest Neighbors, and Decision Tree, using 5-fold cross-validation. Results showed that the proposed deep neural network achieved higher accuracy (79.78%) and precision (76.32%) than the classical models, demonstrating the effectiveness of feature–image fusion for predicting paddy seed viability.

Korean melon (Cucumis melo L.) is an economically important fruit crop in Korea. Hanging cultivation is being introduced to improve space utilization and labor efficiency; however, narrow aisles and cluttered backgrounds challenge autonomous robot operation. This study proposes a hierarchical control framework for a quadruped robot designed to harvest Korean melons in hanging-cultivation greenhouses. Given that conventional ground cultivation remains predominant for Korean melon greenhouses in Korea, the platform is designed with extensibility to ground-cultivation environments. A Behavior Tree orchestrates five modes – navigation, standby, harvesting, post-harvest verification, and manual operation – while a Blackboard shares state and perception outputs across modules. SegFormer-based semantic segmentation estimates walkable regions and extracts a centerline trajectory for heading alignment. During navigation, YOLOv11 detects harvestable fruit from side-camera images. Harvesting is triggered when a melon is detected within a predefined harvestable region, and navigation is resumed after non-detection or verified completion. Hysteresis filtering suppresses mode chattering caused by detection flicker. YOLOv11 achieved a mAP@0.5 of 0.96, and stop-and-mode-transition trials achieved 100% success across ten laboratory tests and five on-site field trials. These results demonstrate that the proposed hierarchical framework enables reliable autonomous harvesting of Korean melons in hanging-cultivation greenhouses using a quadruped robot.

Conventional maize cob phenotyping relies on manual caliper measurement with cylindrical assumption or on 2D RGB imaging, introducing bias for curved, tapered or irregular specimens. This study presents a low-cost depth sensing pipeline that reconstructs complete 3D cob surfaces and extracts continuous ring-wise diameter profiles from principal axis extent, while orthogonal slicing yields local diameter for computing maximum or mean diameter, taper, surface area and volume through true 3D geometry. These 3D traits were benchmarked against (i) caliper-based cylindrical formulas and (ii) a 2D RGB projection baseline with segmentation masking. Validation against manual measurements and repeated scans demonstrates that the 3D methods closely approximate ground truth while substantially reducing error for bent or strongly tapered cobs contexts where 2D projections and geometric simplifications prove inadequate. These findings underscore the practical value of depth-derived diameter profiling and 3D-based volumetric and surface area estimation as an innovative, scalable tool for maize cob phenotyping

Vertical farming is widely recognized as a promising future agricultural model due to its high efficiency in space and resource utilization. Recently, artificial intelligence based robots have been increasingly introduced to establish efficient crop production systems, wherein automated crop growth monitoring is an essential component for optimal facility management. This study proposes a visual classification model for lettuce growth stages in a controlled vertical farm using morphological characteristics and machine learning. Top-down RGB-D images of lettuce were acquired from post-sowing to harvest. A YOLO model was initially applied to localize individual plants. Subsequently, automated image processing techniques were utilized to extract key morphological features, specifically top projected area, cross-sectional area and maximum plant height. These quantitative features trained ML models to classify plants into two stages: GROWTH, and READY. Experimental results using an indoor vertical farm dataset demonstrated that Random Forest classifier achieved the highest classification accuracy of 95.64%, outperforming other models. These findings suggest this AI-driven phenotyping approach can be effectively integrated into robotic systems for automated harvest decision making.