Bridging technology and ecology: enhancing applicability of deep learning and UAV-based flower recognition

Area of Science:

• Ecology
• Computer Science
• Remote Sensing

Background:

• Insect biomass decline poses a significant ecological threat to biodiversity and ecosystem function.
• Effective monitoring of pollinator habitats, particularly floral resources, is vital for conservation strategies.
• Integrating advanced technologies like drones and deep learning can enhance ecological research capabilities.

Purpose of the Study:

• To evaluate the practical application of drone and deep learning technologies for monitoring pollinator habitats.
• To simplify the use of object detection models for flower recognition in ecological research.
• To provide a guide for biologists on applying flower recognition to Unmanned Aerial Vehicle (UAV) imagery.

Main Methods:

• Updated an object detection toolbox to TensorFlow 2 for improved performance and compatibility.
• Tested three object detection models (Faster R-CNN, SSD, EfficientDet) on UAV imagery datasets of flower-rich grasslands.
• Developed a practical guide for applying flower recognition techniques to UAV data.

Main Results:

• Faster Region-based Convolutional Neural Network (Faster R-CNN) achieved the highest overall performance with 89.9% precision and 89% recall.
• EfficientDet offered the lowest model complexity, balancing efficiency and detection performance, though with lower precision than Faster R-CNN.
• The study demonstrated the potential of deep learning models for automated flower detection in UAV imagery.

Conclusions:

• Drone and deep learning technologies offer a promising approach for monitoring pollinator habitats and floral resources.
• Faster R-CNN and EfficientDet are viable models for flower recognition in ecological studies, with trade-offs in performance and complexity.
• Further research is needed to address challenges like dense vegetation and environmental variability in flower detection.