Flower Abundance Maps from Drone Imagery Using Deep Learning

An automated deep learning pipeline generates spatially explicit flower abundance maps from drone images, meeting or exceeding the accuracy of manual counting for some species. The Faster R-CNN model was trained and evaluated on data from five flights across two grassland sites. Precision and recall for some flower types are close to or higher than 90%, though performance varies by species.

Use Cases

• Automating flower abundance surveys in grasslands based on drone imagery.
• Training object detection models for other aerial monitoring tasks based on the described pipeline.
• Studying the impact of phenology on automated species detection based on described appearance changes.
• Assessing the data requirements for reliable detection of small floral features based on described limitations.

Strengths

• Precision and recall for some flower types are close to or higher than 90%.
• Method meets or exceeds the accuracy of manual counting extrapolation.
• Pipeline is described as applicable to any sort of aerial object detection problem.

Limitations

• Some flowers are detected poorly due to lack of training data, phenological changes, or small size.
• Row count, file formats, and column-level documentation are unknown.
• Last update date is unknown; freshness unverified.

Provenance

Source

Johannes Gallmann, ETH Zurich

Collection Method

Drone imagery analyzed with a Faster R-CNN deep learning object detection approach.

Geography

Two grassland sites (specific locations unknown).