Conference Agenda

Variable-depth seeding (VDS) offers potential to improve maize yield under heterogeneous field conditions, yet limited research has explored its practical implementation. This study proposes a two-stage frame-work that integrates machine learning–based depth optimisation with prescription generation and field-scale validation. In the modelling stage, soil physico-chemical properties collected using an online multi-sensor platform were combined with multi-temporal vegetation indices to train maize yield prediction models. Four machine learning algorithms—support vector regression (SVR), random forest (RF), multilayer perceptron (MLP) and extreme gradient boosting (XGBoost)—were evaluated, and optimal seeding depths (OSDs) were obtained by reverse-solving the trained models to identify the depth associated with maximum predicted yield. The resulting OSD maps were implemented in 2025 using an electro-depth-control planter. Results showed that MLP and SVR achieved the highest predictive accuracy (R² = 0.74–0.78), outperform-ing RF and XGBoost (R² = 0.66–0.75). Field evaluation demonstrated that VDS improved yield over uniform-depth seeding, with a mean gain of 0.46 t ha⁻¹ in Groot. In Kleine, VDS 35 mm and 75 mm increased yield, although the reduced performance of VDS 55 mm lowered the mean yield by 0.15 t ha⁻¹. Economic analysis showed a net benefit of 37.2 € ha⁻¹, supporting the economic feasibility of VDS.

Taiwan Native Chicken (TNC) is a highly valued poultry variety in the domestic food supply. TNC is characterized by a longer growth cycle. Thus, breeding TNC to have shorter growth cycles while maintaining its flavor is a critical challenge. Conventional breeding relies on manual observation, which is labor-intensive and subjective. This study developed a system that automatically quantifying traits of TNC for breeding. The system comprised a video acquisition device and a multi-modal deep learning pipeline. When a TNC traversed a capture zone, video recording was triggered. Subsequently, images in recorded videos were identified. Comb and shank of the TNC were detected and segmented, respectively, using YOLOv11 and SAM. Three key features — comb area, shank width, and comb redness ratio — were quantified. In parallel, six keypoints of TNC in the videos were identified using DeepLabCut with a ResNet-50 backbone. With the keypoints, four gait features characterizing limb oscillation, bilateral knee joint angles, and body posture were derived. A one-class SVM was trained to identify anomalies using the traits as inputs. The trained YOLOv11 achieved an mAP@0.5 of 0.98. The trained DeepLabCut achieved a test error of 12.59 pixels. Experimental results confirm that this pipeline enhances selection accuracy over manual methods.

Eggs are an essential staple commodity whose prices and production are influenced by demand–supply interactions and exhibit non-linear volatility. Fluctuations in egg markets affect consumers’ living costs and producers’ strategic decisions. Moreover, price and production time series are subject to seasonality and external shocks, leading to high volatility and temporal dependence across multiple horizons. To address these challenges, this study proposes a Transformer-based multi-horizon forecasting framework that maintains high interpretability. This study analyzes monthly egg data in Taiwan from 2012 to 2023. The model uses historical prices, regional production data, and lagged exogenous variables. Data from 2012-2022 are used for training, and 2023 is used for evaluation. Different combinations of input window lengths and forecasting horizons are examined through rolling forecasts to assess generalization. Experimental results show that the interaction between input window length and forecasting horizon strongly affects predictive performance. Both price and production achieve the best results with a 9-month input window and a 1-month horizon, yielding average MAPEs of 2.10% and 3.88%, respectively. The framework remains robust under volatile market conditions in 2023, delivering stable and actionable forecasts. Future work will analyze model attention weights across forecasting horizons to support practical decision-making and policy applications.

Phalaenopsis plays a pivotal role in the global ornamentals industry, with Taiwan supplying approximately one-third of the world’s market demand. To maintain high-end competitiveness and rapidly adapt to shifting consumer preferences while reducing breeding costs, transitioning toward data-driven precision breeding is essential. A critical prerequisite for this transition is precise phenotyping, specifically the automated identification and characterization of floral organs. This study developed a computer vision pipeline that integrates vision foundation models (VFMs) for the high-precision segmentation of Phalaenopsis floral organs, providing a robust foundation toward the automated quantification of phenotypic patterns. First, a YOLO11n model was trained to detect flower instances, which were subsequently isolated using the SAM2. To achieve organ-level granularity, a DINOv3 with a linear head was fine-tuned on a dataset of 12 manually annotated images to segment five key floral organs: petals, dorsal sepals, lateral sepals, lip and column. Experimental results evaluated on a test dataset demonstrated that the fine-tuned DINOv3 achieved an mIoU of 93.16%. These results confirm that fine-tuning VFMs achieves exceptional performance even with limited labeled data. This research established a scalable digital framework that paves the way for automated pattern quantification, facilitating the transition toward data-driven precision breeding in the orchid industry.

Crop production underpins food security and price stability. Conventionally, yield estimation relied on labor-intensive manual surveys. While advances in imaging technology have enabled image collection via motorcycle-mounted cameras (e.g., GoPro), the subsequent identification process has remained manual and time-consuming. Furthermore, the GoPro datasets are often disorganized, containing anomalies such as non-agricultural elements and low-quality images. These anomalies need to be removed to ensure accurate crop identification. Therefore, this study developed a fully automated pipeline that (i) filters out anomaly images and (ii) classifies 26 crop types within agricultural survey datasets. The proposed framework consisted of four integrated components: (a) image center cropping algorithm (ICCA), (b) non-crop image filter (NCIF), (c) non-target crop image filter (NTCIF), and (d) target crop classification model (TCCM). The ICCA eliminates peripheral noise by focusing exclusively on the central region of the images. The NCIF leveraged Language Segment Anything to filter out non-agricultural backgrounds, such as roads, sky, and infrastructure. The NTCIF employed a Vision Transformer based model with Mahalanobis distance to identify and reject out-of-distribution species. The TCCM performs the final classification of the 26 specific crop types. By filtering complex environmental outliers, the pipeline improves 26-crop classification accuracy, enabling reliable and automated large-scale monitoring.

Swine production constitutes a significant segment of global livestock output. In this sector, gilts play a crucial role in production efficiency. Leg health is a key indicator associated with their economic value and productive lifespan. Conventionally, leg health is assessed by manually observing gait and examining leg structure. However, this approach is time-consuming, labor-intensive, and subject to biases due to observer experience. To address these challenges, this study developed an automatic computer vision approach to detect lameness in pigs. Videos of walking pigs were recorded. A framework based on DeepLabCut and EfficientNet-B0 was developed to detect 16 key anatomical points in video frames. Keyframes were subsequently identified for gait feature extraction. Features included stride time, stance time, and stride length. Finally, lameness was identified using a support vector machine (SVM). A commercial electronic pressure mat system was used to validate the video-based approach. High correlation coefficients of 0.84 in stride time and 0.85 in stance time indicated strong agreement between the measurements. The trained SVM achieved an accuracy of 91.0%. The proposed research provides an automatic and affordable approach to assess pig gait in the pig farming industry.