Real-time biodiversity analysis using deep-learning algorithms on mobile robotic platforms.

Area of Science:

• Ecology
• Robotics
• Artificial Intelligence

Background:

• Global biodiversity decline necessitates advanced conservation strategies.
• Current ecological monitoring methods like capture-recapture are labor-intensive and limited in spatiotemporal scale.
• Stationary sensors provide localized data, lacking comprehensive ecosystem insights.

Purpose of the Study:

• To develop a real-time biodiversity monitoring methodology using deep learning (DL) on mobile robots.
• To overcome limitations of traditional ecological surveys and stationary sensors.
• To enable effective biodiversity quantification and population dynamics investigation.

Main Methods:

• Utilized state-of-the-art deep learning algorithms for real-time species identification on mobile robot payloads.
• Trained DL models achieving high mean average precision (mAP) and fast inference times.
• Developed and tested an experimental payload for field surveys, including online and offline data collection.

Main Results:

• Deep learning algorithms achieved a mean average precision (mAP) of 90.51%.
• Average inference time for DL models was 67.62 milliseconds.
• Field surveys validated the proposed methodology for species identification.

Conclusions:

• The proposed mobile robot-DL system offers a feasible solution for real-time biodiversity monitoring.
• This approach facilitates quick and effective biodiversity surveys, aiding conservation efforts.
• The methodology can be extended for geo-localisation of flora and fauna in diverse ecosystems.