Conference Agenda

This study examines the acceptance of methods aimed at reducing pesticide use in sugar beet cultivation across Bavaria, with a focus on agricultural contractor services combining mechanical weed control and band spraying with modern digital technologies. An online survey conducted in winter 2024/25 yielded 635 valid responses from sugar beet growers, representing approximately 9% of the Bavarian sugar beet farmers and covering about 15% of the cultivation area. The survey assessed current practices, attitudes toward mechanical weed control, and willingness-to-pay for the combined hoeing and band spraying intercompany contractor service using the van-Westendorp price sensitivity approach. Willingness-to-pay for the combined service is substantially below the calculated costs and even lower than the estimated in-house costs of the widely used broadcast spraying method. Regression analysis reveals that greater weed tolerance and favorable regional conditions increase the likelihood of adopting the service, whereas extensive current use of mechanical control decreases it. General skepticism regarding the efficacy of the contractor service remains high, resulting in low purchase probabilities and suggesting a limited market potential. Tailored regional strategies and supportive measures are recommended to enhance the adoption of pesticide-reducing methods in sugar beet farming in Bavaria.

Balancing crop yield and resource efficiency is a critical challenge in modern greenhouse management. Conventional models struggle with this complexity, as mechanistic approaches face parameter uncertainty while data-driven methods lack physical interpretability. To address this, this study propose a hybrid modeling framework to simulate the coupled dynamics of greenhouse microclimate, crop yield, and energy consumption. This approach embeds neural networks into the governing differential equations that describe the system states, constructing a hybrid derivative structure comprising a physical mechanistic term and a neural network correction term. By integrating a differentiable ordinary differential equation solver, the end-to-end joint optimization of physical parameters and network weights is achieved. The framework was robustly validated across high-tech experimental and low-tech commercial greenhouses cultivating sweet pepper, cucumber, and tomato. The results demonstrated that the hybrid model significantly enhanced microclimate prediction accuracy. Compared to the uncalibrated baseline model, the rRMSE for air temperature, relative humidity, and CO2 concentration were reduced by 2.4%, 3.4%, and 7.9%, respectively. Furthermore, the model accurately tracked crop fresh weight accumulation and energy consumption, achieving coefficients of determination above 0.95 for both yield and energy predictions. Ultimately, this framework offers a robust tool for optimizing greenhouse design, environmental control, and crop management.

Acute heat stress poses a significant challenge to global swine production by reducing growth and compromising animal welfare. This study investigates porcine drinking behavior, automatically quantified in the source dataset using a YOLOv11 computer vision pipeline (Bonneau de Beaufort et al., 2024), as a non-invasive digital biomarker for heat stress detection, circadian analysis, and short-term prediction. Nineteen finishing pigs (Large White × Landrace × Piétrain) were exposed to paired six-day thermoneutral (22°C) and heat stress (32°C) conditions, generating 357,120 pig-minute records evaluated under leave-one-pig-out cross-validation (LOPO-CV). A six-hour morning window demonstrated strong detection sensitivity, with drinking duration increasing from 0.350 to 2.621 s/min (7.49-fold; Cohen's d = 2.48, p < 0.0001). Nocturnal drinking proportion increased from 24.28% to 43.39% (+19.11 percentage points; Cohen's d = 1.94, p < 0.0001), indicating strategic behavioral thermoregulation, with K-means clustering identifying eight vulnerable and eleven resilient phenotypes. Day 2 of heat stress strongly predicted Days 2–6 drinking trajectories (LOPO-CV R² = 0.815, RMSE = 0.483 s/min), while Day 1 provided earlier but moderate predictive capability (LOPO-CV R² = 0.269). A candidate alerting threshold of 2.6 s/min in the six-hour morning window is proposed to support targeted cooling interventions and proactive precision livestock management.

Automated pig tracking in precision livestock farming relies on ear tags, which are frequently obscured by mud, blur, variable lighting, and severe occlusions. Traditional machine vision models required massive, manually annotated datasets to generalize under such challenging conditions, making deployment expensive and time-consuming. This study developed a Vision Transformer (ViT) framework for robust alphanumeric ear tag recognition that eliminates the need for manually labeled real-world images. The proposed and baseline models were trained using an online dynamic data generation strategy. Four-character tags were procedurally generated and subjected to intense, randomized spatial and optical distortions to simulate harsh farm conditions. These synthetic tags were then dynamically superimposed onto diverse background images during the training loop. The proposed model achieved an exact match accuracy (4/4 characters) of 81.11% on a test set of 577 real-world images, with 95.84% of cases matching at least three characters. Result showed that misclassifications primarily occurred between visually similar characters, including S and 6, C and 0, and 7 and T. Future work will focus on the design of automatic error-correcting codes and further computational efficiency optimization for edge deployment.

The livestock sector is a major contributor to global GHG emissions yet remains poorly constrained due to heterogeneous management practices and absent point-source reporting. Accurate, asset-level cattle mapping is a critical missing link for transitioning from coarse statistics to high-resolution GHG inventories. We present a dual-track monitoring framework leveraging GeoAI and multi-sensor remote sensing to quantify emissions from intensive and extensive cattle systems. The first track targets intensive systems (CAFOs) using Earth Index to identify confined feeding infrastructure. Central to this is the Bovine Slurry Index (BSI), a Sentinel-2-derived spectral index that isolates organic slurry signatures within operation footprints. Applied across the United States, China, Australia, Brazil, and Argentina, BSI demonstrates that manure ponds—a dominant CH₄ source—share a ubiquitous spectral footprint across regional climates. The second track integrates the FAO Gridded Livestock of the World with high-resolution pasture maps to spatially constrain cattle to realistic grazing lands, correcting misallocation common in national reporting. Combining both tracks, we apply IPCC 2019 Tier 2 methodologies to generate a gridded CH₄ and N₂O dataset stratified by climate zone, productivity class, and management practice. Validated against FAOSTAT totals, this framework provides scalable architecture for asset-level monitoring within the Climate TRACE coalition.