Research on weed identification in soybean fields based on the lightweight segmentation model DCSAnet

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Weeds compete with crops for essential resources, impacting agricultural productivity.
• Modern agriculture increasingly utilizes autonomous systems like robots and drones for efficient weeding and herbicide application.
• The effectiveness of these mobile weeding systems is critically dependent on accurate weed detection capabilities.

Purpose of the Study:

• To introduce DCSAnet, a novel lightweight weed segmentation network designed for enhanced performance on mobile agricultural weeding devices.
• To improve the weed detection accuracy of autonomous weeding systems.

Main Methods:

• Developed DCSAnet, a lightweight network model featuring an encoder-decoder structure and a novel DCA module for feature extraction.
• The DCA module integrates asymmetric and depthwise separable convolutions with channel shuffling, based on MobileNetV3's reverse residual structure.
• Employed a feature fusion strategy in the decoding stage to guide low-dimensional feature aggregation using high-dimensional feature maps, minimizing feature loss.

Main Results:

• DCSAnet achieved a Mean Intersection over Union (MIoU) of 85.95% on a custom soybean field weed dataset.
• The model has a small parameter count of 0.57 million, making it suitable for resource-constrained mobile devices.
• DCSAnet demonstrated superior segmentation accuracy compared to other lightweight networks evaluated.

Conclusions:

• The proposed DCSAnet model offers a highly effective and efficient solution for weed segmentation in agricultural applications.
• Its lightweight design and high accuracy make it well-suited for integration into mobile weeding equipment, thereby improving weed management strategies.