Ash dieback disease poses a severe threat to European ash trees, necessitating improved monitoring and management. However, datasets for training computer vision models for automated ash diabeck disease detection remain limited. To address this, our study investigates a practical computer vision approach to ash dieback detection, using limited real leaflet data augmented by a conditional generative adversarial network (cGAN). A two-phase cGAN training strategy enabled the production of synthetic leaflet images that capture ash-specific features. We test our synthetic data generation on a range of tasks, including classification with models like ResNet and ResNeXt, as well as object detection using YOLO. Results show our synthetic augmentation improves model performance across all tasks. We propose two distinct frameworks to support surveys through semantic segmentation and enable automated data collection for further research. Overall, our approach considers cGANs to enrich limited domain-specific datasets and improve model accuracy across diverse vision tasks, and offers headway in applying learning frameworks to enhance biodiversity conservation over current methods.

Ash dieback, caused by the fungal pathogen Hymenoscyphus fraxineus, is devastating ash tree populations across U.K. and Europe, with projections indicating that up to 80% of ash trees may die as a result of the disease. The extensive loss of this keystone species threatens biodiversity and may lead to significant habitat degradation. Since no cure exists, early detection and removal of infected trees are critical to slowing the spread of the disease. Traditional identification methods rely on visual assessments of canopy loss, which are inefficient and impractical for large-scale monitoring. Leveraging advancements in computer vision and deep learning, our key objective is to develop a tool to detect ash dieback symptoms at the leaf level, classifying leaves into three categories: healthy, early-stage infection, and mid-stage infection. Since there is no known available dataset for ash dieback at the leaf level, we generated a new synthetic dataset and trained a YOLOv5 single-stage object detection model. The final model achieves mean Average Precision (mAP) scores of above 90% for each category. Evaluations on real ash tree leaf footage captured using uncrewed aerial vehicles (UAVs) show strong alignment between the model’s detections and expert annotations. Our tool demonstrates the potential of integrating advanced computer vision techniques into tree health monitoring platforms. In the near future, this can provide conservationists and researchers with a novel, efficient means of early disease identification.

Aerial manipulators have the unique ability to cover wide-spread areas within a single mission, making them ideal for the transport and placement of sensors required to build an instrumented environment. Recent work in the field has focused on controllers for aerial interaction that account for compliance during contact-based tasks, omitting integration concerns that are critical to an automated solution. Furthermore, state-of-the-art flying base manipulators are often mechanically and computationally complex, reducing their endurance. Within this work, we present an interactive framework for autonomous sensor placement that incorporates both mechanical and software based compliance, optimised for use on a simple coplanar quadrotor. Under appropriate actuation and perception constraints, we detail the development of a control, perception, and motion planning strategy to enable sensor placement that relies solely on onboard computation and sensing, thus presenting a fully contained and accessible sensor placement approach capable of robust interaction with the environment. An extended finite-state machine is developed to facilitate automated mission planning. Extensive flight experiments are performed to validate the effectiveness of each sub-system, as well as the integrated solution. Experiments result in trajectory tracking errors under 10 mm as well as onboard mass estimation errors under 0.7% for sensors of various weights. A statistical analysis of 162 flight experiments shows the proposed framework's ability to autonomously place sensors within 10 cm of the target with a success rate of 93.8% and 95% confidence interval of (89%, 97%), thus confirming the robustness of our approach.1.

Aerial manipulation is a nascent research area that offers major impact for infrastructure monitoring and repair. While several design and control methods have been presented, there is still a need for new mechatronic solutions that are structurally optimised for aerial manipulation tasks. In this paper we present a novel design for a manipulator tailored for aerial applications with a high level of morphological integration with the robot frame. A hybrid system is presented that comprises a 5-bar linkage parallel robot with an additional active joint for the swirling motion about a pivotal point. The design offers an omnidirectional workspace about the aerial vehicle, enhancing the versatility of the aerial system and the tasks that can be accomplished. The mechanical design of the proposed robot, the analysis of the kinematics and the study of the workspace are presented. The novel manipulator represents the first of its kind, enabling aerial interaction with ceilings, curved surfaces and side interaction with facades.