Conference Agenda

Taiwanese native chickens (TNCs) are highly valued. Conventional manual TNCs inspection remains labor-intensive. This study developed an automated system for TNC carcass defect inspection and removal, comprising an image acquisition module, a computer vision pipeline, and a removal module. The computer vision pipeline included a blurred image exclusion algorithm (BIEA), a silky fowl exclusion algorithm (SFEA), a body parts detection model (BPDM), and a body parts defect classification model (BPDCM). The BIEA and SFEA, respectively, excluded blurred images and silky fowl, ensuring that only valid TNC carcasses were processed by BPDM and BPDCM. The BPDM detected four body parts –– wings, legs, breast, and back –– of chicken carcasses. The BPDCM classified these body parts as defective or sound. The BIEA and SFEA achieved 100% accuracy in filtering out invalid images. The trained BPDM achieved a mean average precision of 99.1% on detecting body parts in carcass images. The trained BPDCM achieved an accuracy of 96.5% on identifying defective body parts. On-site validation demonstrated an overall system accuracy of 86.5% with an average processing time of 608.2ms per carcass. These results demonstrate the system's potential to automate carcass classification and removal, addressing labor shortages in TNC slaughterhouses.

Monitoring of fruit purée consistency is still largely based on manual Bostwick tests, which are time consuming and do not allow continuous control. In this study, data from an industrial apple purée processing line were used to predict consistency directly from inline sensors. A dataset of 524 samples was collected from an operating production line, including pressure drop, temperature, flow parameters, and physicochemical descriptors of the apples.

Three modelling strategies were evaluated, starting from a physics-based baseline using pressure drop and progressively integrating process and fruit variables. Generalized Linear Models, Gradient Boosting Machines, and Deep Learning architectures were compared. The best performance was achieved by the Gradient Boosting Machine, reaching R² = 0.78 and MAPE = 9.2%, demonstrating robust prediction of Bostwick consistency without direct laboratory measurements.

Results highlight pressure difference and thermal conditions as dominant predictors, while ripeness-related attributes significantly improved model generalization. The proposed approach enables real-time estimation of rheological behavior through indirect sensing, supporting adaptive process control and reduced manual testing. Beyond the specific case study, the methodology offers a scalable solution for AI-based monitoring of complex food flows in agro-industrial processing systems.

Honey fraud and origin mislabeling are long-standing problems for smallholder beekeepers in Brazil, pushing producer prices down and weakening buyer confidence across the supply chain. As part of the Semear Digital initiative, we developed and evaluated a digital certification system for honey cooperatives built on open web identity standards.

The approach relies on the W3C Verifiable Credentials Data Model 1.1 combined with the decentralized identifier DID:Web method, which anchors issuer identity to a cooperative-managed web domain. Each issued certificate captures batch-level data including geographic origin, quality parameters and organic certifications, along with photographic field evidence uploaded at the time of issuance. Ed25519 signatures bind all this information to issuing cooperative´s public key, enabling third-party verification over standard HTTPS without requiring centralized intermediaries.

A fully functional application was built and tested end-to-end, covering cooperative-managed beekeeper identity registration, credential issuance and QR-based delivery to a Progressive Web App wallet that runs on any smartphone without installation. Beekeepers retain full ownership of their credentials stored locally on their own devices, while public verification by buyers or auditors requires no account, no dedicated app and no technical background.

Results suggest that DID:Web infrastructure can offer a low-cost, interoperable traceability layer for agricultural cooperatives.

The morphological assessment of grapevine (Vitis vinifera L.) bunches remains a challenging task requiring skilled staff, particularly when evaluating ambiguous traits such as bunch compactness. Unlike objective descriptors (e.g., berry number and weight), compactness is traditionally assessed through visual inspection, which lacks sensitivity and reproducibility. This study integrates observations from three experimental trials aimed at investigating bunch morphology through digital twin reconstruction. Three-dimensional models were generated using photogrammetry and artificial intelligence algorithms, enabling a comprehensive morphometric characterization of bunch morphology. The derived traits were evaluated against objective variables associated with compactness, specifically: (i) the relationship between Pinot Gris clone morphology and susceptibility to bunch rot; (ii) the relationship between Moscato Giallo bunch morphology and drying performance. Morphological analysis was conducted extracting 28 two-dimensional and 36 three-dimensional descriptors. Feature selection was performed through analysis of variance (ANOVA). Estimated empty volume, vertical section area, and the areas of the upper and lower horizontal sections emerged as the most informative variables for describing bunch compactness across the three trials. On the other hand, artificial intelligence reduces the general effort for the three-dimensional reproduction generation despite photogrammetry. Further comparative analyses across diverse post-harvest contexts will help validate and generalize this approach.

Precise rice disease monitoring is vital for food security. While large multi-modal models (LMMs) excel in general understanding, they struggle with fine-grained agricultural pathology. Initial evaluations show baseline models like vanilla LISA achieve an F1 of only 0.3890 and a gIoU of 0.2240, revealing significant limitations in specialized tasks. To bridge this gap, we propose LISA-DAA, an interactive assistant for rice disease diagnosis. We designed a three-stage progressive knowledge transfer paradigm to overcome data scarcity: (1) general visual feature initialization; (2) plant pathology domain adaptation via dynamic weight-tilting; and (3) architecture enhancement using a Domain-Adaptive Attention (DAA) module to refine lesion boundaries. Results demonstrate the superior efficiency of LISA-DAA. Compared to the baseline, the model achieved a gIoU of 0.7240 (a 223% increase), an F1 score of 0.8953, and a Precision of 0.8450. Notably, the Rice-IoU reached 0.7120, marking a qualitative leap in segmentation accuracy. Furthermore, the mAP rose to 0.9012, confirming high stability in complex field environments. This research transforms LISA into an interactive diagnostic assistant with expertlevel reasoning, providing a robust algorithmic foundation for large-scale, full- growth-cycle rice disease monitoring

Extreme weather conditions and changes in rainfall patterns have been a major issue affecting cashew nut yields. An adequate forecast of climate change has a direct impact on production and the cultivation areas. Therefore, an optimized data-driven machine learning cashew nut yield prediction model is proposed. Feature engineering was performed to improve model correctness, computational efficiency, reduce overfitting, and enhance model interpretability.

The study employed Principal Components analysis, which revealed that area, production, annual rainfall, fertilizer, and pesticide are the major contributing factors for efficient cashew yield prediction. An Artificial Intelligence-enabled mobile application called AI-CashewYeild was developed to address the challenges in cashew nut yield prediction due to irregularities in weather and climate changes. Machine learning algorithms such as Light Gradient Boosting Machine Regressor, Extra Tree Regressor, Support Vector Regressor, Random Forest Regressor, and Multi-Layer Perceptron Regressor were implemented to build the models. The performance metrics, including R², Mean Absolute Error, Mean Squared Error, and Root Mean Squared Error, were used for models’ evaluation. The performance results showed that MLP Regressor is the golden model with 0.998 R², 0.01 MAE, 0.008 MSE, and 0.02 RMSE. The AI-CashewYeild app can be used to monitor and forecast weather conditions for higher yield..