Flower Mapping in Grasslands With Drones and Deep Learning.

Area of Science:

• Ecology
• Botany
• Computer Science

Background:

• Manual assessment of grassland flower abundance is labor-intensive and time-consuming.
• Accurate monitoring of plant species distribution and abundance is crucial for ecological studies.

Purpose of the Study:

• To develop and evaluate an automated deep learning approach for determining flower abundance in grasslands using drone-based aerial imagery.
• To compare the accuracy and efficiency of the automated method against traditional manual assessment techniques.

Main Methods:

• Utilized a deep learning object detection model (Faster R-CNN) trained on drone-based aerial images from multiple flights and sites.
• Developed a pipeline for generating spatially explicit maps of flower abundance.
• Evaluated model performance using precision and recall metrics for various flowering plant species.

Main Results:

• The deep learning network successfully identified and classified individual flowers, achieving high precision and recall (often >90%) for certain species.
• Generated accurate, spatially explicit maps of flower abundance, outperforming manual extrapolation methods in efficiency.
• Detection performance varied by species, with challenges noted for small or phenologically variable flowers due to limited training data.

Conclusions:

• The developed automated method offers a less labor-intensive and highly accurate alternative for mapping grassland flower abundance.
• Future improvements can be achieved by collecting more diverse training data to enhance predictions for challenging species.
• The established pipeline is adaptable for various aerial object detection applications beyond floral abundance mapping.