Segmenting vegetation from UAV images via spectral reconstruction in complex field environments

Area of Science:

• Remote Sensing
• Computer Vision
• Agricultural Science

Background:

• Field vegetation segmentation is crucial for monitoring but challenged by complex environments.
• Current spectral sensing and deep learning methods are hindered by high equipment costs and limited, labor-intensive annotated datasets.
• Existing approaches struggle with the cost and time demands of data acquisition and annotation.

Purpose of the Study:

• To develop a cost-effective and efficient weakly supervised method for field vegetation segmentation.
• To overcome limitations of high equipment costs and manual data annotation in remote sensing.
• To improve vegetation monitoring and analysis in complex field conditions.

Main Methods:

• Proposed SRCNet and SRANet models for multispectral image reconstruction using spectral reconstruction (SR) techniques.
• Integrated spectral reconstruction with vegetation index (VI) principles to enhance vegetation information.
• Employed a weakly supervised adaptation strategy for segmentation without manual labeling.

Main Results:

• Achieved a Mean Intersection over Union (MIoU) of 0.853 on real field datasets, outperforming existing methods.
• Successfully reconstructed multispectral images, reducing the need for expensive spectral data collection.
• Demonstrated effective field vegetation segmentation using a weakly supervised approach.

Conclusions:

• The proposed weakly supervised method offers an efficient and accurate solution for field vegetation segmentation.
• Spectral reconstruction and VI aggregation effectively address data acquisition costs and annotation burdens.
• The open-sourced dataset and code will advance research in remote sensing for agriculture.