Conference Agenda

Accurate estrus detection is essential for improving reproductive efficiency in cattle, but conventional methods remain labor-intensive and difficult to sustain under farm conditions. This study developed a non-invasive vision-based deep learning framework for monitoring estrus-related behaviors in Hanwoo cattle using progressive dataset refinement, targeted augmentation, and comparative lightweight object detectors. The initial dataset was highly imbalanced and the rare estrus-related behavior class was underrepresented. After iterative cleaning and relabeling, a refined benchmark was established in which the estrus-related interaction category included mounting, sniffing, and chin-resting behaviors. Using a partial-context interaction labeling strategy, YOLOv5n, YOLOv8n, and YOLO11n were compared. Current pilot results showed that YOLOv8n achieved the highest precision and the best refined-stage mAP@0.5 for estrus_interaction, whereas YOLO11n achieved the strongest recall. Targeted augmentation further improved minority-class sensitivity, indicating that rare-class performance depends not only on detector choice but also on biologically meaningful data refinement. These findings indicate that the main challenge is robust detection of rare estrus-related interactions under class imbalance, occlusion, and overlap, and that recall is particularly important for practical estrus monitoring because missed events are more critical than a modest increase in false positives.

This study proposes a four-stage experimental framework to systematically evaluate how image restoration affects object detection performance in pig-farming scenes with inter-animal overlap and structural occlusions. Rather than emphasizing perceptual restoration quality alone, we quantitatively assess whether restoration leads to measurable gains in downstream detection. In Stage 1, detection models (YOLO11, Faster R-CNN, and RT-DETR) are trained on 1,776 images annotated into four classes (PIG, OC_PIG, IB_PIG, OCIB_PIG). The model with the highest validation F1-score is selected to generate 640×640 cropped images (S2) from predicted bounding boxes, reducing irrelevant background. In Stage 2, an amodal segmentation dataset is constructed from these crops. Images are manually labeled with two-class masks (modal visible region and amodal restoration region), and synthetic occlusions are added via randomly inserted boxes to expand occlusion diversity without additional labeling cost. In Stage 3, three amodal segmentation models (SAMBA, AISFormer, YOLO-based segmentation) are benchmarked, and the best model generates restoration masks for inpainting. Inpainting models are selected based on LPIPS, PSNR, and SSIM. Finally, Stage 4 fixes detector weights and compares detection results before (S2) and after restoration (S4), analyzing metric changes, class transitions, and false prediction reductions.

Reliable pest monitoring forms the foundation of integrated pest management. We present an automated workflow that detects, classifies, and counts pest insects on commercially available yellow sticky traps integrated into a custom-developed monitoring system (Smart Checkpot).

Our approach integrates classical computer vision techniques with neural net-work–based data modelling. We implemented image processing routines using Halcon (www.mvtec.com) and trained neural networks with Membrain (www.membrain-nn.de). The investigations used yellow sticky cards, placed in insect breeding facilities, which were subsequently photographed. After image acquisition, we manually annotated the data for training and validation. The image processing pipeline applies colour space transformations and threshold-based fil-tering to segment individual insects. Each insect was described by 21 geometric features capturing relevant morphological properties.

We trained backpropagation-based classifiers on a dataset containing three insect species, 100 sticky cards, approximately 600 annotated insects, and 125 unseen test samples. Sensitivity analyses showed that a statistically motivated reduction of the feature space significantly improved classification accuracy. The optimized model achieved an overall accuracy of about 85%.

By combining rule-based image processing with feature reduction and neural networks, the system delivers a robust solution for automated pest monitoring and currently outperforms the YOLO algorithm.

Rising global temperatures challenge Precision Livestock Farming (PLF), yet characterizing behavioral responses to heat stress remains difficult due to noise in automated monitoring. This study quantifies behavioral plasticity and rest fragmentation in growing pigs (N=19) under thermoneutral (22°C) and heat stress (32°C) conditions. A discrete-state Hidden Markov Model (HMM) was applied to continuous computer vision time-series data, inferring four latent states: Sustained Rest, Restless/Transition, Hydration/Thermoregulation, and Feeding. Results revealed a dual reorganization of behavioral dynamics under heat stress. Macro-temporally, the behavioral distribution contracted, with a 46% reduction in feeding occupancy and an 87% increase in Hydration/Thermoregulatory behaviors. Micro-temporally, Sustained Rest became highly fragmented; heat stress reduced mean resting bout duration by 44.6% (16.97 to 9.39 min), indicating disrupted rest despite high lying prevalence. These findings demonstrate that while pigs display behavioral plasticity to cope with thermal challenges, this adaptation compromises rest stability. This probabilistic framework captures multiscale dynamics undetectable by conventional time budgets, providing a robust tool for assessing livestock thermal comfort.

Vision-based reinforcement learning (RL) offers a route to automated crop control in controlled environment agriculture (CEA). Yet progress is hampered by the scarcity of time-series crop images matched to corresponding measurements and environment records. We present LettuceVisSim, a lettuce growth simulator that generates multimodal time-series datasets pairing shoot dry weight with top view images at a user configurable sampling interval. This simulator contains a process-based crop growth model and two components. Firstly, a canopy layout algorithm maps individual shoot dry weight to a canopy layout representation, which captures canopy scale ground coverage dynamics. Secondly, a Unity rendering engine converts the layout representation into visually plausible top view RGB images and binary segmentation masks. Validation results showed that the canopy layout representation closely reproduces canopy development observed in measured images, both qualitatively and quantitatively, throughout cultivation under dynamic environment and management. Efficiency evaluation indicated that simulation speed depends on image resolution and sampling frequency and is comparable to existing simulators under low image resolutions commonly used in vision-based RL. The simulator enabled a proof-of-concept study, demonstrating that RL can learn effective lighting control directly from images. Overall, LettuceVisSim alleviates multimodal data scarcity and supports data-driven control research in CEA.

Adequate description of ambient thermal conditions benefits precision agriculture as it supports research on heat stress, thermal comfort zones and thermoregulation. However, thermal analysis often requires the assistance of computer-aided designs (CAD) models linked to a sensing system or numerical solutions of computational fluid dynamics (CFD) models assisted with data of temperature or heat fluxes. Either CAD or CFD models rely on thermal boundary conditions in order to achieve comprehensive thermodynamic assessments comprising spatio-temporal heat exchanges due to sinks and sources observed within complex geometries. In this research is proposed the use of current technological advances associated to images and machine learning (ML) methods to reduce difficulties defining thermal boundary conditions. The approach consisted of capturing paired visual and thermal images of the same region of interest for subsequence processing with ML methods. The visual image was used to reconstruct three-dimensional (3D) geometric representations whereas thermal images were projected onto those 3D geometric representations. This approach provided an effective technological resource requiring minimal requirements for calibration and allowing minimization of complexity to explain heat exchanges. Thus becoming a choice for enabling real-time thermodynamics, enhanced energy assessment and optimization.