Lightweight weed segmentation model based on UAV multispectral imagery

Weeds have restricted winter wheat growth and yield. Conventional pesticide spraying can trigger waste, residues, and ecological issues. Among them, UAV multispectral remote sensing can be expected to efficiently monitor weeds in precise field management. However, existing deep learning models cannot well balance segmentation accuracy and lightweight deployment: DeepLabv3+ is parameter-heavy for edge devices, and lightweight models such as MobileNetV2 perform poorly under complex field conditions. This study aims to propose the lightweight model (Weed-DeepLab) on the DeepLabv3+ framework with UAV multispectral data. A tradeoff was also obtained for high accuracy and lightweight architecture. An attentional fusion inverted residual (AFIR) module was constructed using an AFIR-Mobile backbone network to replace the original Xception structure. The new backbone network fully met the lightweight requirements and multi-scale feature representation. A wavelet texture augmentor (WTA) module with wavelet transform convolution (WTConv) was designed to extract fine texture features and spatial distribution. A fine-grained channel attention (FCA) mechanism was introduced after the Atrous spatial pyramid pooling (ASPP) module. Spectral differences were fully exploited to improve multi-scale feature selection and inter-class discrimination. Systematic experiments were conducted on the same dataset with eight mainstream models, including AFFormer-base, Rolling-UNet-M, PIDNet-M, HRNet, I2UNet-M, Swin-UNet, and DeepLabv3+. The experimental results demonstrated that Weed-DeepLab steadily outperformed all models in terms of the various evaluation indicators. The mean intersection over union (mIoU), mean precision (mPre), mean F1-score (mF1), and accuracy (Acc) of the improved model reached 81.08%, 88.06%, 88.83%, and 96.61%, respectively. Four evaluation metrics increased by 5.31, 5.96, 3.87, and 1.26 percentage points, respectively, compared with DeepLabv3+. The parameter counts of the model decreased significantly from 54.71 to 6.02 M. Compared with Swin-UNet, the four metrics improved by 1.30, 1.73, 0.91, and 0.30 percentage points, respectively. The parameter volume was reduced by 85.45%. The mIoU and mF1 values increased by 2.09 and 1.36 percentage points, respectively, compared with the lightweight model (Rolling-UNet-M). The parameter quantity decreased by 1.08 M, whereas the inference speed increased by 154.23 frames per second. Ablation experiments further verified the synergistic effects of AFIR, WTA, and FCA modules. The AFIR module was used to optimize the preservation of crop and weed boundary information. The WTA module was enhanced to extract field texture features. The FCA module was improved to fuse spectral and multi-scale spatial features. Qualitative results showed that the Weed-DeepLab maintained full edge features of target objects. Low misclassification rates were also obtained under weed-free, sparse-weed, and dense-weed scenarios, indicating better environmental robustness. Edge-device validation tests were performed on the Jetson Xavier NX platform. The model realized an inference speed of 24.16 frames per second when processing 128×128 five-channel multispectral images without additional acceleration frameworks. The speed fully met the real-time requirements of field weed identification. In conclusion, Weed-DeepLab effectively relieved the contradiction between segmentation accuracy and lightweight deployment. Accurate and rapid weed detection was realized under three typical field scenarios of winter wheat at the tillering stage. Favorable segmentation, high accuracy, and a compact parameter scale were suitable for the reliable edge deployment. The finding can provide reliable technical support for intelligent weed recognition, targeted field control, and lightweight model deployment on resource-constrained agricultural equipment in winter wheat fields.