SpikeSegNet-a deep learning approach utilizing encoder-decoder network with hourglass for spike segmentation and...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• High-throughput, non-destructive phenotyping is crucial for identifying superior crop traits and understanding genetic factors (QTLs).
• Accurate detection and counting of wheat spikes are essential for phenomics, breeding programs, and precision agriculture.
• Digital image analysis and machine learning are key technologies for non-destructive plant analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for accurate, non-destructive detection and counting of wheat spikes.
• To enhance the efficiency of phenotyping for large-scale germplasm and breeding populations.
• To provide a tool for precision agriculture applications requiring critical stage monitoring.

Main Methods:

• A deep learning network, SpikeSegNet, was developed, integrating Local Patch extraction Network (LPNet) and Global Mask refinement Network (GMRNet).
• SpikeSegNet utilizes computer vision and object detection techniques on digital RGB images of wheat plants.
• ImageJ's "analyse particles" function was used for spike counting on the segmented images generated by SpikeSegNet.

Main Results:

• SpikeSegNet achieved high precision (99.93%), accuracy (99.91%), and robustness (99.91% F1 score) in spike segmentation.
• The spike counting method demonstrated average precision of 99%, accuracy of 95%, and robustness of 97%.
• The approach maintained segmentation performance even with illuminated image datasets, indicating robustness.

Conclusions:

• SpikeSegNet is an effective and robust deep learning approach for identifying and counting wheat spikes non-destructively.
• This method represents a significant advancement in non-destructive, high-throughput phenotyping of wheat.
• Accurate spike detection and counting using SpikeSegNet can contribute to improved crop yield prediction and management.