Deep Learning-Based Semantic Segmentation of Paddy Field Weeds: A Comparative Study of Multi-Model Performance and Perspectives for Smart Agriculture

Abstract

This paper presents a comparative analysis of six semantic segmentation models for rice seedling and weed identification: SegNet, EfficientNet-SegNet, MobileNet-SegNet, DeepLabV3-ResNet50, DeepLabV3-ResNet101, and SegFormer. Motivated by the need for automated solutions in precision agriculture to reduce labor, improve crop management, and reduce herbicide usage, this study addresses the challenges of identifying morphologically similar and spatially mixed weeds and rice seedlings. The aim was to evaluate the segmentation performance and deployment feasibility under class imbalance and complex background conditions typical of paddy field imagery. The metrics include pixel accuracy, mean Intersection over Union (mIoU), per-class IoU, model complexity, and inference efficiency. SegNet achieved the highest mIoU (0.7204) and outperformed CED-Net (0.7105). MobileNet-SegNet balances the accuracy and speed, whereas SegFormer delivers competitive accuracy with the lowest parameter count and FLOPs. Paired t-tests (p < 0.0001) confirmed the statistical significance of the performance differences, offering practical insights for selecting models in resource-constrained agricultural settings.