Conference Agenda

Cultivation in controlled environment agriculture (CEA) systems is an innovative strategy to optimize water use and enhance crop production in limited spaces. This study evaluated the effects of bioinputs and different artificial light intensities on the biophysical and physiological responses of strawberry (Fragaria vesca sp.) grown under controlled environment agriculture. A completely randomized design was adopted in a 2 × 4 factorial scheme, with two light intensities (1296 and 648 µmol m⁻² s⁻¹) and four bacterial treatments (Bacillus amyloliquefaciens, Priestia aryabhattai, a combination of both species, and a non-inoculated control). Physiological variables included photosynthetically active radiation (PAR), photosynthetic rate, stomatal conductance, transpiration, vapor pressure deficit (VPD), and photosynthetic pigments, while morphological traits related to plant growth were also assessed. The results revealed significant interactions between light intensity and bacterial inoculation for key physiological parameters, particularly photosynthetic rate and pigment concentration. Higher light intensity promoted superior physiological performance, especially when associated with Priestia aryabhattai and single-species inoculation. Bioinputs enhanced photosynthetic efficiency under high irradiance, whereas combined inoculation showed limited benefits for some pigments. These findings demonstrate the potential of integrating optimized light management and microbial bioinputs to improve strawberry performance in CEA systems, reinforcing this approach as a sustainable and resource-efficient production strategy.

Controlled Environment Agriculture (CEA) is gaining importance due to its high resource-use efficiency, particularly in irrigation water management. Increasing climate variability further emphasizes the need for protected and climate-resilient production systems. Advancing CEA requires research infrastructures capable of providing precise, stable, and reproducible growing conditions for testing cultivation strategies and technical solutions.

We describe the design, construction, and operational performance of a low-cost indoor CEA platform (16 m² floor area, 3.8 m height) developed for research and pilot-scale experimentation. The system integrates automated climate control based on temperature, relative humidity, and CO₂ sensors; timer-based irrigation; LED lighting; in-substrate EC monitoring; and thermal imaging for non-invasive plant status assessment. The platform was designed to ensure spatially and temporally uniform environmental conditions.

A seven-month pilot cultivation was conducted using ‘Favori’ everbearing strawberries grown on two three-tier vertical racks. Environmental monitoring showed highly uniform conditions, with mean within-chamber temperature fluctuation of 0.048 °C and relative humidity fluctuation of 0.72%. Average berry yield reached approximately 975 g per plant. The developed CEA platform provides a scalable and flexible infrastructure for precision indoor cultivation research and validation of advanced climate control strategies.

Protected horticulture requires accurate, spatially resolved prediction of greenhouse microclimate to optimize crop performance and energy use. However, conventional building energy simulation tools typically rely on lumped air models that neglect local gradients, while computational fluid dynamics is computationally expensive for real-time applications. This study develops a Physics-Informed Neural Network (PINN) framework to predict 2-D temperature distribution and airflow velocity inside a single-span greenhouse. The PINN model predicts temperature and air velocity by embedding the transient heat transfer equation with boundary and initial conditions, and enforcing incompressible Navier–Stokes with continuity constraints. Models were trained using limited sensor measurements and collocation points, implemented in PyTorch with tanh activation and Adam optimization. Results show progressive convergence with training, producing smooth, physically consistent thermal and velocity fields without meshing or iterative solvers. Temperature prediction achieved high agreement with experiments, outperforming comparable ANN-based approaches. The airflow model captured realistic ventilation-driven circulation patterns, though it tended to underestimate magnitudes due to sparse measurements and 2-D simplification. Overall, the proposed PINN offers a rapid, data-efficient alternative for greenhouse digital twins and real-time climate and energy control.

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2025-00560277).

The paper reports on the development and testing of an automated digital trap for Drosophila suzukii to overcome limitations of conventional monitoring systems, which rely on periodic collection of insects captured in liquid or chemical attractants and labor-consuming laboratory identification and counting. The proposed device is designed to reduce field inspection costs, technician workload, and improve identification accuracy through remote image acquisition. The trap was conceived using robust, low-cost hardware and open-source software solutions. An innovative design was implemented to allow imaging of insects prior to their fall into the attractant reservoir. An enclosed transparent chamber was created between the entry holes and the attractant access point, ensuring a constant focal distance from the camera. Insect entry interrupts a laser beam detected by photocells, triggering both LED illumination and image acquisition (photo or video mode). The system is powered by a photovoltaic panel with battery storage and enables local data storage and remote transmission. The prototype, entirely 3D-printed, was tested under controlled laboratory conditions and in a greenhouse. Although capture rates were lower than those of traditional traps, temporal trends were consistent. Results confirm the technical feasibility of an automated digital monitoring system and highlight its potential for real-time, intelligent pest management applications.

The development of resilient, data-driven crop production systems requires controlled-environment phenotyping platforms capable of generating reproducible, high-quality data under precisely manipulated conditions. Reliable analysis of plant responses to environmental variables and validation of smart control strategies demand growth environments where temperature, humidity, CO₂, light, irrigation, and nutrient supply are accurately regulated and continuously monitored.

Within the DiAGRI initiative, Metropolia University of Applied Sciences is establishing an advanced controlled-environment phenotyping platform at UrbanFarmLab. The facility integrates climate-controlled growth rooms with programmable logic controllers (PLCs), distributed I/O systems, and edge-computing infrastructure. Custom-designed mechanical structures and automated rigging systems enable repeatable image acquisition and precise sensor positioning. A dedicated phenotyping station combines RGB imaging, 3D scanning, hyperspectral imaging, chlorophyll fluorescence imaging, thermal cameras, and substrate sensors. Environmental and crop data are collected via networked IoT systems and integrated into a cloud-connected data lake supporting AI-based multimodal data fusion. The automation architecture and motion sequences were validated through virtual commissioning and kinematic simulation prior to deployment.

The platform supports development of digital twins, predictive yield models, and closed-loop climate optimization strategies, enabling standardized phenotyping workflows and scalable AI-driven solutions for sustainable controlled-environment agriculture.

Pest outbreaks cause substantial agricultural losses worldwide and pose a growing threat to global food security. According to the Food and Agriculture Organization, nearly 40% of global crop production is lost to pests, while food demand is projected to increase by approximately 50% by 2050. In addition, climate change is altering pest distribution and population dynamics, making conventional manual monitoring methods inefficient and insufficient for timely pest management. This study proposes a deep learning-based object detection approach using three species-specific models for species-level monitoring of Spodoptera litura, Spodoptera exigua, and Riptortus clavatus from images collected by automated trap systems. Field-collected trap images were used for training and validation, and images without ground-truth boxes were also included during training to reduce false detections caused by non-target objects in species-specific traps. Three separate detection models were developed, and their performance was evaluated using Precision, Recall, mAP@50, and mAP@75 metrics.

Digital Twin (DT) technology offers significant potential for precision agriculture by enabling real-time crop monitoring and data-driven field management. This study presents an ongoing effort to develop and implement an online DT workflow for an artichoke open field operating under real agricultural conditions in Sardinia (Italy). Vegetative and geometrical crop data were acquired through a multi-sensor approach integrating Unmanned Aerial Systems (UAS), Unmanned Ground Vehicles (UGV), satellite imagery, soil sensors, and operator field surveys for pest and disease detection. Acquired data were structured into standardised, date- and geo-stamped datasets, augmented with AI-based computer vision inference (e.g., plant-level disease symptom classification), and fed into the DT pipeline. The prototype DT enables interactive visualisation of individual plants via a 3D model (e.g., GLB/glTF format), providing plant-level status information to support site-specific management decisions. However, development has exposed critical open challenges: the lack of full autonomy in unmanned vehicle operation, insufficient automation in data acquisition and processing, non-standardised dataset formatting, inadequate server communication pipelines, and suboptimal 3D model storage and streaming efficiency for seamless online rendering. This work discusses these limitations and outlines the key development priorities required to advance the system toward operational deployment in real agricultural scenarios.

Healing agriculture has emerged as a sustainable approach to enhancing psychological well-being and rural resilience. However, systematic engineering frameworks for utilizing bioacoustic resources, particularly sound-producing insects, remain limited. A Mel-spectrogram is a time–frequency representation that converts acoustic signals into spectral patterns suitable for deep learning analysis. This study provides a foundational investigation for developing an insect acoustic–based digital healing platform and evaluates digitalization and classification technologies for insect acoustic data. A total of 5,674 Mel-spectrogram images were generated from WAV recordings of 19 insect species segmented into 0.5-second intervals. Mel-spectrograms were computed using 128 Mel filter banks while preserving the original sampling rate. A hybrid noise-reduction pipeline combining band-pass filtering and spectral thresholding was implemented to suppress background noise. Subsequently, 1,089 contaminated samples (19.19%) were removed through automated screening and manual verification, resulting in 4,585 high-quality images. All images were resized to 256×256 pixels and normalized before training, and the dataset was divided into training, validation, and test sets (7:2:1). An EfficientNetB0-based transfer learning model with dropout regularization was employed for classification. The proposed framework achieved 99.79% test accuracy with a weighted F1-score of 0.9979, demonstrating its potential for scalable digital healing agriculture platforms.

Crop phenotyping involves the systematic measurement of plant traits, such as size, leaf area, architecture, and biomass, to characterize plant responses to environmental conditions and management practices. It plays a central role in modern agriculture by supporting crop improvement and enhancing resource-use efficiency. Conventional phenotyping methods are typically labor-intensive and time-consuming, particularly under field conditions, where manual measurements can be inconsistent and prone to error. Advances in imaging and sensing technologies have therefore promoted the development of non-destructive, high-throughput phenotyping approaches, with three-dimensional (3D) imaging emerging as a promising solution.

This study compares a conventional two-dimensional (2D) RGB workflow with a 3D RGB-D approach based on a low-cost sensor for the non-contact estimation of lettuce biomass from a top-view perspective. Two 3D processing strategies were evaluated: voxelization of individual plant point clouds and voxelization combined with a reconstruction algorithm designed to compensate for occluded structures. Results highlight the potential of low-cost 3D imaging combined with dedicated algorithms to provide accurate and scalable solutions for crop phenotyping.

STEM (Satellite Tree Extraction & Metrics) is an operational tool developed by the Biomass Laboratory at Università Politecnica delle Marche to automatically detect reserve trees (“matricine”) in coppice stands using high‑resolution satellite imagery. The methodology combines digital image processing, morphological analysis, LiDAR data, and semi‑automated Python procedures to build a replicable pipeline that extracts spatial metrics for monitoring and verifying harvesting plans. The system uses a 50‑cm panchromatic WorldView‑2 image, contextualized with a 1‑m LiDAR‑derived DTM, and was tested on three coppice sites in Basilicata, Italy. The workflow includes pre‑processing and spatial alignment of the area, followed by binarization with morphological filtering to reduce noise and separate objects. A watershed transform then segments individual crowns, after which polygons and oriented bounding boxes (OBB) are generated for each object. From these, centroids, orientation, mean slope (from the DTM), and inter‑crown distances are derived. OBBs provide the basis for future object‑detection models (e.g., YOLO), while the metrics support silvicultural assessments. The prototype confirms the technical feasibility of recognizing residual trees in coppice stands, with practice‑consistent results: average spacing of ~11 m, typically larger on steeper slopes. STEM is a concrete step toward operational monitoring systems grounded in objective, transparent, and verifiable data.

Methane is a major air pollutant emitted from pig housing and can adversely affect the overall barn environment; however, its temporal variability and key drivers remain insufficiently quantified under commercial conditions. This study aimed to monitor CH₄ concentrations using continuous sensor-based measurements and to identify influential environmental and management factors using machine-learning models. Data quality was ensured by detecting outliers with a Hampel filter and handling missing values via hybrid imputation. Random Forest, LightGBM, and histogram-based gradient boosting (HistGB) models were developed and compared. HistGB provided the best predictive performance (R² = 0.83) and the lowest RMSE. Model interpretation indicated that feed intake, slurry pH, and temperature were the principal determinants of CH₄ emissions, whereas ventilation rate and relative humidity predominantly explained variability in ammonia (NH₃). These findings demonstrate that robust preprocessing combined with gradient-boosting approaches can reliably characterise gas concentration dynamics and prioritise controllable drivers, thereby supporting improved gas mitigation and management strategies in intensive pig production systems.

Recent advances in artificial-intelligence have significantly improved the efficiency and precision of disease detection in both plants and animals. Early and accurate identification of diseases is critical for preventing severe losses, and AI-based systems offer fast and reliable solutions compared to traditional manual inspections. This study presents an experimental analysis using image recognition and machine learning techniques to enhance disease detection performance. Two approaches are evaluated.

The first approach uses convolutional neural networks (CNNs) to enhance PlantVillage images of apples and corn. Features extracted through CNNs are classified using a Bayesian support vector machine (SVM), achieving high accuracy and precision. This method demonstrates strong potential for preventing large‑scale agricultural losses by enabling timely disease intervention.

The second approach extracts texture and color features from dataset images using histograms of oriented gradients (HoG). These features are combined into hybrid representations incorporating color, texture, and depth information. A random forest classifier is applied to the optimized feature set selected using binary particle swarm optimization, ensuring effective classification with minimal feature redundancy.

Evaluation results confirm the effectiveness of both strategies. The study highlights how AI and computer vision can significantly improve early disease detection, support better crop management, and reduce economic-losses for farmers.

Precision Agriculture represents a systemic transformation of the agricultural production model, integrating digital technologies into agronomic practice to enable site-specific, data-driven, and temporally dynamic management with site-specific resource allocation.. This study aims to characterize soil and crop variability in an experimental olive grove at Azienda Xiggiari, La Pergola (Paceco, TP) through multisensor integration: multispectral and thermal UAV, airborne LiDAR and Veris iScan surveys, validated through field-based vegetation sampling. Radiometrically calibrated orthomosaics were produced, together with NDVI maps (separated for canopy and soil), thermal maps (ΔT canopy–soil), canopy area maps, soil maps (ECa, Red/NIR reflectances, NIR/Red ratios) and voxelized 3D LiDAR models. A GEOBIA approach enabled accurate canopy segmentation; comparison of tree‑by‑tree analysis and a 10×10 m grid revealed consistent patterns (low‑vigour margins, central areas of medium–high vigour). Canopy NDVI (NDVIcs) and canopy area were strongly correlated (R² ≈ 0.81–0.84). Calibrated LiDAR, validated with cubic samples, estimated leaf number and leaf area with R² up to 0.96, and canopy volume showed a high correlation with NDVIcs (R² = 0.89). Thermal maps and Veris surveys highlighted microclimatic and pedological gradients consistent with vegetative responses. Multisensor integration provides a robust basis for zoning and variable‑rate management; further validation with yield data is required for operational adoption.