Conference Agenda

Precision orchard management requires accurate and consistent tree-level information across time and space. This study presents an integrated aerial–ground sensing framework that combines Agroview, a cloud-based AI-driven remote sensing platform, with Agrosense v2, a real-time ground-based orchard sensing system, to deliver scalable, multi-attribute tree-level monitoring. The objective was to evaluate whether fusing aerial analytics with in-field AI perception improves reliability and actionable decision support.

Agroview analyzed multi-date aerial imagery to extract plant-level metrics, including inventory, canopy height, canopy area, and leaf density, and assessed temporal consistency across two citrus blocks. Agrosense v2 integrated tree detection and counting, canopy density classification, tree height estimation, and detection-based fruit counting on both tree and ground into a unified pipeline with geo-referenced aggregation. Field validation was conducted on 3,696 citrus trees across commercial and research orchards.

Inventory variation from Agroview among several data collection dates was below 3%, and canopy height coefficients of variation were below 9% for most trees. Agrosense v2 achieved tree counting accuracies up to 97.8%, canopy density classification up to 96.8%, and fruit detection mAP@0.5 up to 0.837. The results demonstrate that integrating aerial consistency with ground-level detail enables robust, scalable orchard intelligence for precision interventions and data-driven management.

Volumetric units (wet basins, ponds, and clarifiers for at-grade applications or as tanks/tunnels below-grade in surface-constrained areas) are widely implemented for water reuse after hydrologic control and constituent load reduction. Herein, a computational model and results thereof are illustrated to optimize cost/benefit of volumetric units in Florida which leads the USA in such reuse. The model combines computational fluid dynamics (CFD) and artificial intelligence (AI) simulations over a wide range of volumetric unit configurations, loadings, hydrodynamics and particulate matter (PM) granulometry. A novel augmentation of CFD with AI models is developed and trained to create surrogate clarification models. This CFD-AI platform facilitates optimization of existing volumetric unit retrofits to minimize cost for a required level of treated (clarified) urban drainage reuse. Results with CFD-AI benchmarking indicate: (a) residence time (RT) models are not accurate or generalizable for treated reuse, (b) RT models are agnostic to geometrics, hydrodynamics, PM granulometry; and do not reproduce treatment needed for reuse, (c) trained AI models provide high predictive capability (± 15%) as surrogates for CFD simulations; yielding rapid computations. CFD-AI demonstrate that simply enlarging volumetric units based on RT presumptive guidance results in exponential cost increases, irrespective of infrastructure adjacency conflicts and reuse benefits.

Mechanical hoeing can loosen topsoil and increase erosion risk on sloped sites. This study assessed the technical feasibility of site-specific, autonomous mechanical weed control in maize. A complete workflow was implemented on a 0.15-ha field, including drone-based georeferenced imaging, AI-based weed detection, aggregation into a 3 × 3 m grid to define treatment zones, and execution with the Robotti 150D (AgroIntelli) and a Chopstar 3-60 hoe (Einböck).

High-resolution imagery (3 mm px⁻¹) at BBCH 12 was analysed to detect weeds, resulting in 1,786 polygons. Weed coverage was calculated per grid cell. For demonstration, a 6.2% coverage threshold defined treatment zones; this value was exemplary and not agronomically optimised. The prescription map was transferred to the vehicle planning platform, enabling selective hoeing. In total, 0.08 ha (53%) of the field was treated.

Results demonstrate the technical feasibility of site-specific mechanical weed control using autonomous systems. However, routine application is currently limited. Effects on erosion risk, labour, and fuel consumption depend strongly on the size and distribution of treated patches. Efficiency gains are highest when few, spatially distinct weed patches require treatment. Further research should quantify labour and energy impacts.

Sunburn in apple production occurs when the fruit surface temperature exceeds 45 °C due to high air temperatures and/or solar radiation. This causes discoloration and surface damage and has been reported to reduce yields by 10-50%, with incidence expected to increase as climate change worsens. Although evaporative cooling by misting is a common mitigation strategy, uneven droplet distribution can create gaps in cooling. In this study, we propose a system that can acquire environmental data from an apple orchard and overlay them onto a three-dimensional (3D) representation of the orchard. A 3D reconstruction of part of an apple orchard was generated from images processed with the MUSt3R algorithm, and a point cloud was exported. Environmental data were obtained from a CSV file with data from the Japan Meteorological Agency website, which can also be used to read data directly from edge devices, and visualized in real time in a web browser by overlaying them onto the 3D model using the PyVista module. System performance, including reconstruction quality and memory usage, was evaluated. The results suggest that this system has potential to be further developed into a digital twin for apple orchards.

Re-identifying individual cows over time is essential for precision
dairy farm management, enabling long-term monitoring tasks such as health
assessment and behavior tracking. A common practice in re-identification is to
construct a gallery offline with labeled identities; however, building such a
gallery is costly, and a static gallery cannot incorporate newly appearing cows.
This limits practical open-world deployment, where identities are not fixed and
previously unseen individuals may appear over time. This study presents an
autonomous cow face re-identification and online enrollment pipeline using farm
CCTV streams. A YOLO-based detector localizes cow faces, which are
temporally associated into tracklets for robust identity evidence accumulation. A
ResNet50 embedding network trained with ArcFace loss on a labeled source
domain is adapted to the target farm via clustering-guided pseudo-labeling. For
open-world deployment, the system incrementally builds an adaptive gallery
through staged enrollment and self-adaptive prototype management, regulating
gallery growth while limiting error propagation. Enrollment is governed by
calibrated similarity thresholds and confidence margins, and their impact on
open-world discovery is discussed. In an online evaluation, the proposed system
achieved strong track-level retrieval performance (mAP = 95.11%) while
maintaining a representative adaptive gallery. Overall, the results demonstrate
that the approach reduces manual labeling and gallery initialization effort,
supporting deployable open-world cow face re-identification for scalable long-
term cow monitoring.