.png){kind=icon}
.png){kind=icon}
Return
Hybrid CNN-LSTM-channel attention model for multi-class seedling classification using RGB images
.png){kind=icon}
PRE
AI
E
F
A
S
DOI：10.1007/s00521-026-12374-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The accurate identification of plant seedlings remains one of the most difficult tasks in precision agriculture, where timely crop monitoring and informed decisions on weed control are crucial for sustainable farming. The automatic detection of seedlings is much more challenging because of the strong visual similarity between crops and weeds in their early developmental stages. A new hybrid deep learning architecture is proposed by combining Convectional Neural Networks (CNNs), Long Short-Term Memory (LSTM) units, and a channel-wise attention module that can address such critical challenges as inter-class visual similarity, dataset class imbalance, and environmental variability. To avoid problems related to dataset class imbalance—where the risk of poor generalization from a minority class may be large—the class-weighted loss functions ensure that each category of a seedling contributes equally during the network training process. On the V2 Plant Seedlings dataset, the proposed CNN-LSTM-Attention framework reached a test accuracy of 93.2%, outperforming strong baselines of ResNet18, EfficientNet-B0, and CNN-Gated Recurrent Units (GRU). Moreover, the channel-wise attention in the proposed architecture enhances model interpretability in highlighting biologically-significant plant regions, such as the leaf margin and midrib structures. In the real-time application of automated agricultural systems, where precision and efficiency are both key issues, this paper presents an effective approach toward overcoming these challenges.
Keywords:
Precision agriculture
Plant seedlings
Hybrid model
Image classification
CNN-LSTM-Attention
Journal
IF:
4.5
Papers:
729
Citations:
3.2W
Organization
Cited Papers