Smart weeding with less data: New AI model learns to spot rare weeds in record time

The model, called the Few-Shot Enhanced Attention (FSEA) network, incorporates plant-specific features—such as color cues and morphology—into its learning process. By integrating domain knowledge with advanced attention mechanisms, FSEA enables rapid and accurate adaptation to unfamiliar weeds in diverse field environments.

Weeds severely reduce crop yield and quality, and excessive herbicide use threatens both ecosystems and human health. Deep learning has revolutionized plant detection, but its data-driven nature demands vast, balanced datasets that are nearly impossible to obtain under field conditions. Agricultural images often feature occlusions, variable lighting, and an uneven distribution of weed species, limiting the generalizability of current models. FSOD offers a potential solution by enabling fast model adaptation from limited data. However, existing FSOD models lack domain-specific optimization for agricultural conditions, particularly when weeds overlap or vary greatly in morphology. Based on these challenges, researchers developed the FSEA network for efficient few-shot weed detection.

A study (DOI: 10.1016/j.plaphe.2025.100086) published in Plant Phenomics on 5 July 2025 by Jingyao Gai’s team, Guangxi University,  reduces the need for large, time-consuming datasets and paves the way for intelligent, eco-friendly weed management systems suitable for precision agriculture and sustainable crop production.

In this study, the Few-Shot Enhanced Attention (FSEA) network was evaluated against six state-of-the-art few-shot detectors (TFA, FSCE, Meta R-CNN, Meta-DETR, DCFS, and DiGEO) and a traditional detector (YOLOv7) to assess its adaptability to new weed species under limited training data. After 40 epochs of fine-tuning, FSEA demonstrated superior performance, achieving an all-class mean average precision (mAP) of 0.416 and a novel-class mAP of 0.346, outperforming all baseline methods. In contrast, fine-tuning-based models such as TFA and FSCE exhibited poor adaptation and severe base–novel trade-offs, while meta-learning-based approaches (Meta R-CNN, Meta-DETR) achieved more balanced but lower overall accuracy. Feature enhancement-based models (DCFS and DiGEO) improved feature discrimination but struggled with occlusion and small-object detection. The general-purpose YOLOv7 model overfitted under few-shot conditions, confirming its limited suitability for this task. Ablation experiments further validated each FSEA module: the feature fusion module increased base and novel mAP by 0.081 and 0.061, respectively, by focusing attention on green vegetation features; the feature enhancement module raised mAP by 0.105 and 0.044 by better capturing plant morphology; and the repulsion loss improved occlusion handling by an additional 0.024 and 0.014. Qualitative analyses confirmed that FSEA maintained robust detection across plant sizes and occlusion levels while achieving real-time inference at 32 frames per second. These results demonstrate that by integrating vegetation-specific color, morphology, and occlusion priors, FSEA effectively overcomes the data scarcity challenges in agricultural weed detection, ensuring both accuracy and efficiency in practical field environments.

This research offers a powerful solution for modern precision agriculture, enabling weeding robots and vision-based monitoring systems to adapt quickly to new environments without extensive retraining. The FSEA model’s integration of plant-specific priors reduces data requirements and enhances accuracy under real-world field conditions. Beyond weed detection, its methodology can extend to rare plant identification, pest monitoring, and early crop disease diagnosis. The open-source release of its dataset and code provides a foundation for further research in agricultural artificial intelligence. By promoting efficient, selective weed control, FSEA supports sustainable farming practices and minimizes reliance on chemical herbicides—advancing both agricultural productivity and environmental conservation.

###

References

DOI

10.1016/j.plaphe.2025.100086

Original URL

https://doi.org/10.1016/j.plaphe.2025.100086

Funding information

This project was funded by National Natural Science Foundation of China, China (Award No.: U23A20330) and Specific Research Project of Guangxi for Research Bases and Talents, China (Award No.: AD22035919). This research is also a product of the Modern Industry School of Subtropical Intelligent Agricultural Machinery and Equipment, Guangxi University, China (Project No. T3010097930).

About Plant Phenomics

Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.