A Parameter-efficient Convolutional Approach for Weed Detection in Multispectral Aerial Imagery

Researchers Leo Thomas Ramos and Angel D. Sappa have introduced FCBNet, a novel computer vision architecture specifically designed for efficient weed segmentation in agricultural aerial imagery. The model combines a fully frozen ConvNeXt backbone with a custom Feature Correction Block (FCB) that uses efficient convolutions for feature refinement, followed by a lightweight decoder. This approach allows FCBNet to maintain high accuracy while dramatically reducing computational requirements.

FCBNet was rigorously evaluated against established models including U-Net, DeepLabV3+, SK-U-Net, SegFormer, and WeedSense on both RGB and multispectral data from the WeedBananaCOD and WeedMap datasets. The model achieved a mean Intersection over Union (mIoU) score exceeding 85%, outperforming all competitors. Most impressively, the frozen backbone strategy reduced trainable parameters by more than 90%, enabling training times of just 0.06 to 0.2 hours while maintaining superior segmentation quality.

The architecture's efficiency stems from its parameter-efficient design, where the pre-trained ConvNeXt backbone remains completely frozen during training, requiring only the lightweight FCB and decoder to be optimized. This approach not only reduces memory requirements but also makes the model particularly suitable for deployment in resource-constrained agricultural settings where computational power may be limited but accurate weed detection is crucial for precision farming applications.