Rodent hole detection in a typical steppe ecosystem using UAS and deep learning.

Area of Science:

• Ecology and Environmental Science
• Remote Sensing and Geospatial Technology
• Artificial Intelligence and Machine Learning

Background:

• Rodent outbreaks pose significant threats to grassland ecosystems, necessitating effective monitoring.
• Traditional rodent damage monitoring methods, such as field surveys, are often costly and labor-intensive.
• Brandt's voles (BV) in Inner Mongolia present a monitoring challenge due to small, seasonally obscured burrow openings.

Purpose of the Study:

• To develop and evaluate a novel framework integrating unmanned aircraft systems (UAS) and deep learning (DL) for efficient Brandt's vole (BV) hole detection.
• To establish the first bi-seasonal UAS image datasets specifically for rodent hole detection.
• To compare the performance of various deep learning object detection models for BV hole identification.

Main Methods:

• A novel framework combining UAS image acquisition and deep learning (DL) models was proposed for BV hole detection.
• Two bi-seasonal UAS image datasets were created for training and validating DL models.
• Six object detection models (Faster R-CNN, R-FCN, Cascade R-CNN, SSD, RetinaNet, YOLOv4) were evaluated based on accuracy, speed, and generalizability.

Main Results:

• Faster R-CNN and YOLOv4 demonstrated the highest accuracy in detecting BV holes.
• SSD and YOLOv4 exhibited the fastest processing speeds.
• Faster R-CNN and YOLOv4 showed the most consistent performance across different seasons.

Conclusions:

• The integration of UAS and DL provides an automatic, accurate, and efficient method for BV hole detection in steppe ecosystems.
• This approach holds significant potential for large-scale, multi-seasonal rodent damage monitoring.
• The developed framework and datasets contribute to advancing ecological monitoring techniques.