Automatedly identify dryland threatened species at large scale by using deep learning

Area of Science:

• Ecology
• Remote Sensing
• Artificial Intelligence

Background:

• Dryland biodiversity is declining rapidly, necessitating advanced tools for conservation.
• Populus euphratica, a threatened species in China's Tarim River Basin, has experienced significant population decline.

Purpose of the Study:

• To evaluate deep learning models for automated, large-scale detection of threatened dryland species using high-resolution satellite imagery.
• To compare the performance of Deeplabv3+, Unet, and Pspnet for identifying Populus euphratica.

Main Methods:

• Utilized high-resolution (0.3 m) unmanned aerial vehicle (UAV) satellite imagery for species identification.
• Applied three deep convolutional neural network models: Deeplabv3+, Unet, and Pspnet.
• Assessed model performance based on varying training dataset sizes and mean pixel accuracy (MPA).

Main Results:

• Deeplabv3+ achieved the best overall performance with 80 training samples, showing MPA between 87.31% and 90.2%.
• Deeplabv3+ outperformed Unet and Pspnet by 3.74% and 11.29% on average, respectively.
• Deeplabv3+ accurately identified Populus euphratica boundaries, even in dense vegetation, with lower pixel uncertainty.

Conclusions:

• A UAV-based deep learning framework provides accurate, high-resolution identification of threatened dryland species in large, inaccessible areas.
• This approach supports rapid and efficient conservation actions for species like Populus euphratica.
• The study highlights the potential of advanced AI for monitoring and conserving vulnerable ecosystems.