Context Weed infestations significantly reduce cotton productivity by competing with crops for resources, leading to significant decreases in yield and fibre quality. Conventional weed identification approaches are time-consuming, labour-intensive and often inaccurate. To tackle these problems, the use of advanced automated techniques is needed to increase yields, to promote sustainable farming practices and to promote precision farming. These challenges create barriers to the achievement of the UN Sustainable Development Goals (SDGs), in particular SDG 2 (zero hunger), SDG 12 (responsible consumption and production) and SDG 15 (life on earth) which require sustainable food systems with responsible agrochemical processing and protection of terrestrial ecosystems. Objective This study develops EFCENet, an explicable attention-aware deep learning (DL) model to achieve accurate and computationally efficient classification of cotton and weed. This model allows for an intelligent and affordable weed monitoring system that minimises herbicide use and is in line with the UN SDGs direct benefits to SDG2 (zero hunger) through improved food security, SDG12 (responsible consumption and production) through reduced agrochemical use and SDG15 (life on earth) through protection of soil health and biodiversity. Methods A comprehensive dataset of 11,232 cotton and weed images has been carefully developed and evaluated. The proposed architecture integrates a reinforced parallel Convolution Block Attention Module (CBAM) with EfficientNet-B1 backbone to enhance the extraction of discriminant information. The reliability of the models was further assessed by comparing them with established convolutional neural network (CNN) architectures, bootstrap statistics and interpretability evaluated by gradient-weighted class activation mapping (Grad-CAM). Results The proposed EFCENet achieves a 99.64% accuracy, 99.65% precision, 99.64% recall and 99.64% F1 scores, while requiring 7.31 million parameters and 1.37 GFLOPs of computing power. A bootstrap analysis confirmed the statistical reliability with a 95% confidence interval of 99.35 to 99.88%. Comparative evaluation demonstrated superior performance compared with EfficientNet-B1(99.35%), MobileNetV2(99.41%), Inception-V3(99.11%), Inception-ResNetV2(98.81%), NasNetMobile(98.10%) and Xception(98.34%). The discriminative capacity of the model was further confirmed by the AUC-ROC values of 0.998 for cotton and 0.997 for weed. Grad-CAM visualisations confirmed the effective location of the discriminant plant regions, which supported the interpretability of the model and the transparency of the decision.. Conclusions Enhanced parallelism mechanism allows for better extraction capabilities, improves classification results while maintaining efficient processing. The proposed EFCENet, achieving 99.64% accuracy with only 7.31 million parameters and 1.37 GFLOPs demonstrates superior performance compared to existing CNN architectures, and offers a reliable, lightweight, and interpretable solution for automated cotton–weed discrimination suitable for real-world precision agriculture deployment. Implications. The explainable attention-based DL model EFCENet enables precise and reliable weed management classification. By achieving 99.64% recall, EFCENet ensures that missed weed detections which cause yield losses of up to 30% in cotton are minimised, directly advancing SDG 2 (Zero Hunger). Its targeted classification capability supports site-specific herbicide application, reducing agrochemical use by an estimated 35–40% of conventional input costs and thereby advancing SDG 12 (Responsible Consumption and Production). Lower chemical loads mitigate soil microbiome disruption and surface-water contamination, contributing measurably to SDG 15 (Life on Land). With a 27.90 MB memory footprint deployable on low-cost mobile hardware, EFCENet reduces labour requirements and improves profitability for smallholder farmers, supporting SDG 1 (No Poverty) and SDG 8 (Decent Work and Economic Growth).

Explainable attention-based deep learning framework for automated cotton weed classification

期刊论文
出版模式：

开放获取
发表日期：

2026年7月16日

查看全文

Ritika,
Savita Kumari Sheoran

摘要

Context

Weed infestations significantly reduce cotton productivity by competing with crops for resources, leading to significant decreases in yield and fibre quality. Conventional weed identification approaches are time-consuming, labour-intensive and often inaccurate. To tackle these problems, the use of advanced automated techniques is needed to increase yields, to promote sustainable farming practices and to promote precision farming. These challenges create barriers to the achievement of the UN Sustainable Development Goals (SDGs), in particular SDG 2 (zero hunger), SDG 12 (responsible consumption and production) and SDG 15 (life on earth) which require sustainable food systems with responsible agrochemical processing and protection of terrestrial ecosystems.

Objective

This study develops EFCENet, an explicable attention-aware deep learning (DL) model to achieve accurate and computationally efficient classification of cotton and weed. This model allows for an intelligent and affordable weed monitoring system that minimises herbicide use and is in line with the UN SDGs direct benefits to SDG2 (zero hunger) through improved food security, SDG12 (responsible consumption and production) through reduced agrochemical use and SDG15 (life on earth) through protection of soil health and biodiversity.

Methods

A comprehensive dataset of 11,232 cotton and weed images has been carefully developed and evaluated. The proposed architecture integrates a reinforced parallel Convolution Block Attention Module (CBAM) with EfficientNet-B1 backbone to enhance the extraction of discriminant information. The reliability of the models was further assessed by comparing them with established convolutional neural network (CNN) architectures, bootstrap statistics and interpretability evaluated by gradient-weighted class activation mapping (Grad-CAM).

Results

The proposed EFCENet achieves a 99.64% accuracy, 99.65% precision, 99.64% recall and 99.64% F1 scores, while requiring 7.31 million parameters and 1.37 GFLOPs of computing power. A bootstrap analysis confirmed the statistical reliability with a 95% confidence interval of 99.35 to 99.88%. Comparative evaluation demonstrated superior performance compared with EfficientNet-B1(99.35%), MobileNetV2(99.41%), Inception-V3(99.11%), Inception-ResNetV2(98.81%), NasNetMobile(98.10%) and Xception(98.34%). The discriminative capacity of the model was further confirmed by the AUC-ROC values of 0.998 for cotton and 0.997 for weed. Grad-CAM visualisations confirmed the effective location of the discriminant plant regions, which supported the interpretability of the model and the transparency of the decision..

Conclusions

Enhanced parallelism mechanism allows for better extraction capabilities, improves classification results while maintaining efficient processing. The proposed EFCENet, achieving 99.64% accuracy with only 7.31 million parameters and 1.37 GFLOPs demonstrates superior performance compared to existing CNN architectures, and offers a reliable, lightweight, and interpretable solution for automated cotton–weed discrimination suitable for real-world precision agriculture deployment.

Implications.

The explainable attention-based DL model EFCENet enables precise and reliable weed management classification. By achieving 99.64% recall, EFCENet ensures that missed weed detections which cause yield losses of up to 30% in cotton are minimised, directly advancing SDG 2 (Zero Hunger). Its targeted classification capability supports site-specific herbicide application, reducing agrochemical use by an estimated 35–40% of conventional input costs and thereby advancing SDG 12 (Responsible Consumption and Production). Lower chemical loads mitigate soil microbiome disruption and surface-water contamination, contributing measurably to SDG 15 (Life on Land). With a 27.90 MB memory footprint deployable on low-cost mobile hardware, EFCENet reduces labour requirements and improves profitability for smallholder farmers, supporting SDG 1 (No Poverty) and SDG 8 (Decent Work and Economic Growth).