DSC-DeepLabv3+: a lightweight semantic segmentation model for weed identification in maize fields.

Introduction: Weeds compete with crops for water, nutrients, and light, negatively impacting maize yield and quality. To enhance weed identification accuracy and meet the requirements of precision agriculture, we propose a lightweight semantic segmentation model named DSC-DeepLabv3+. Methods: MobileNetV2 is adopted as the backbone, and standard convolutions in atrous spatial pyramid pooling (ASPP) and decoder modules are replaced with depthwise separable dilated convolutions (DSDConv), significantly reducing model complexity and improving segmentation efficiency. To capture rich contextual information, strip pooling is incorporated into the ASPP module, forming the strip pooling-atrous spatial pyramid pooling (S-ASPP) structure. In addition, a convolutional block attention module (CBAM) is introduced to refine feature representations, and multi-scale features are further fused using the CBAM-Cascade Feature Fusion (C-CFF) module to improve semantic understanding. Results: Experimental results show that the proposed model reduces the number of parameters from 54.714M to 2.89M and decreases the computational cost from 167.139 GFLOPs to 15.326 GFLOPs, while achieving an inference speed of 42.89 FPS and a mean Intersection over Union (mIoU) of 85.57%. Discussion: These results demonstrate that DSC-DeepLabv3+ strikes an effective balance between accuracy and efficiency, outperforming several classical lightweight models, making it a promising solution for accurate and efficient weed segmentation in agricultural applications.