RCTUnet: a deep learning model for crop-residue-soil image segmentation and crop residue cover extraction

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate crop residue cover (CRC) quantification is vital for conservation tillage evaluation.
• Image segmentation for CRC faces challenges due to subtle visual differences, illumination variations, and shadows.

Purpose of the Study:

• To introduce RCTUnet, a novel deep learning architecture for robust crop-residue-soil segmentation and precise CRC estimation.
• To overcome the limitations of existing methods in segmenting fragmented residue from soil in field imagery.

Main Methods:

• Developed RCTUnet, integrating a ResNet50 backbone, a convolutional block attention module (CBAM), and a transformer-based global context fusion module (GCFM).
• Evaluated RCTUnet on 1220 field-acquired images across four crop rotations.
• Compared RCTUnet's performance against Unet, Unet++, DeepLabV3, SegNet, and FCN.

Main Results:

• RCTUnet achieved higher overall accuracy in crop-residue-soil segmentation compared to traditional models (e.g., 3.24% improvement over Unet).
• Demonstrated superior residue-soil segmentation performance with increased residue recall (e.g., 7.67% increase over Unet).
• Showcased enhanced CRC estimation accuracy with a 45.5% reduction in RMSE compared to Unet (4.875 vs. 8.941).

Conclusions:

• RCTUnet's hybrid approach effectively combines deep features, attention mechanisms, and global context modeling for improved CRC assessment.
• The model offers a robust and reliable tool for automated CRC quantification, advancing in-field agricultural monitoring.
• RCTUnet significantly outperforms existing deep learning models for agricultural image segmentation tasks.