Zhijun Dai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zhijun Dai has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
Heng Zhou
Hunan Engineering and Technology Research Center for Agricultural Big Data Analysis & Decision-making, China
2 shared publicationsQingfang Ba
Hunan Engineering and Technology Research Center for Agricultural Big Data Analysis & Decision-making, China
2 shared publicationsZheming Yuan
Hunan Engineering and Technology Research Center for Agricultural Big Data Analysis & Decision-making, Hunan Agricultural University, China
2 shared publicationsChangcheng Li
Hunan Engineering and Technology Research Center for Agricultural Big Data Analysis & Decision-making, Hunan Agricultural University, China
1 shared publicationDongcheng Li
Hunan Engineering and Technology Research Center for Agricultural Big Data Analysis & Decision-making, Hunan Agricultural University, China
1 shared publication- Dongcheng Li
Hunan Engineering & Technology Research Center for Agricultural Big Data Analysis & Decision-making, Hunan Agricultural University, Changsha 410128, China
1 shared publication
Latest Publications
- Open Access
- 8 August 2025
- Published by Frontiers Media SA in Frontiers in Plant Science
- Vol. 16, 1619365
LiSA-MobileNetV2: an extremely lightweight deep learning model with Swish activation and attention mechanism for accurate rice disease classification
In the context of intelligent agriculture in China, rapid and accurate identification of crop diseases is essential for ensuring food security and improving crop yield. Although lightweight convolutional neural networks (CNNs) are widely adopted for plant disease recognition due to their computational efficiency, they often suffer from limited feature representation and classification accuracy. To address these challenges, we propose LiSA-MobileNetV2, an improved MobileNetV2-based model designed for rice disease classification. First, we restructure the inverted residual module to simplify the network architecture, achieving a test accuracy of 92.32%, representing a 2.41% improvement over the original MovileNetV2 (89.91%). This indicate that a more lightweight network can enhance feature representation in specific disease recognition. Second, integrating the Swish activation function further improves accuracy to 94.04% by enhancing the model’s non-linear feature learning. Finally, the addition of a squeeze-and-excitation attention mechanism raises accuracy to 95.68%, representing a 5.77% improvement over the original model. Importantly, the parameter size and FLOPs are reduced by 74.69% and 48.18%, respectively, maintaining strong computational efficiency. These results demonstrate that combining structural simplification, advanced activation, and efficient attention mechanisms significantly improves CNN performance. LiSA-MobileNetV2 provides a high-accuracy, resource-efficient solution for real-time rice disease detection in smart farming systems.
- 8 August 2025
- Published by Frontiers Media SA in Frontiers in Plant Science
- Vol. 16, 1619365
Open AccessLiSA-MobileNetV2: an extremely lightweight deep learning model with Swish activation and attention mechanism for accurate rice disease classification
In the context of intelligent agriculture in China, rapid and accurate identification of crop diseases is essential for ensuring food security and improving crop yield. Although lightweight convolutional neural networks (CNNs) are widely adopted for plant disease recognition due to their computational efficiency, they often suffer from limited feature representation and classification accuracy. To address these challenges, we propose LiSA-MobileNetV2, an improved MobileNetV2-based model designed for rice disease classification. First, we restructure the inverted residual module to simplify the network architecture, achieving a test accuracy of 92.32%, representing a 2.41% improvement over the original MovileNetV2 (89.91%). This indicate that a more lightweight network can enhance feature representation in specific disease recognition. Second, integrating the Swish activation function further improves accuracy to 94.04% by enhancing the model’s non-linear feature learning. Finally, the addition of a squeeze-and-excitation attention mechanism raises accuracy to 95.68%, representing a 5.77% improvement over the original model. Importantly, the parameter size and FLOPs are reduced by 74.69% and 48.18%, respectively, maintaining strong computational efficiency. These results demonstrate that combining structural simplification, advanced activation, and efficient attention mechanisms significantly improves CNN performance. LiSA-MobileNetV2 provides a high-accuracy, resource-efficient solution for real-time rice disease detection in smart farming systems.
- Open Access
- 18 June 2025
- Published by Frontiers Media SA in Frontiers in Bioinformatics
- Vol. 5, 1607119
Enhancing genomic prediction in Arabidopsis thaliana with optimized SNP subset by leveraging gene ontology priors and bin-based combinatorial optimization
With the rapid development of high-density molecular marker chips and high-throughput sequencing technologies, genomic selection/prediction (GS/GP) has been widely applied in plant breeding. Arabidopsis thaliana, as a common model organism, provides important resources for dissecting genetic variation and evolutionary mechanisms of complex traits. Quantitative traits are typically influenced by multiple minor-effect genes, which are often functionally related and can be enriched within gene ontology (GO) pathways. However, optimizing marker subsets associated with these pathways to enhance GP performance remains challenging. In this study, we propose an improved GS framework called binGO-GS by integrating GO-based biological priors with a novel bin-based combinatorial SNP subset selection strategy. We evaluated the performance of binGO-GS on nine quantitative traits from two A. thaliana datasets, comprising nearly 1,000 samples and over 1.8 million SNPs. Compared with using either the full marker set or randomly selected markers with Genomic BLUP (GBLUP), binGO-GS achieved statistically significant improvements in prediction accuracy across all traits. Similar improvements were observed across six additional regression models when applying binGO-GS instead of the full marker set. Furthermore, the selected markers for identical or similar morphological traits exhibited consistent patterns in quantity and genomic distribution, supporting the polygenic model of complex quantitative traits driven by minor-effect genes. Taken together, binGO-GS offers a powerful and interpretable approach to enhance GS performance, providing a methodological reference for accelerating plant breeding and germplasm innovation.
- 18 June 2025
- Published by Frontiers Media SA in Frontiers in Bioinformatics
- Vol. 5, 1607119
Open AccessEnhancing genomic prediction in Arabidopsis thaliana with optimized SNP subset by leveraging gene ontology priors and bin-based combinatorial optimization
With the rapid development of high-density molecular marker chips and high-throughput sequencing technologies, genomic selection/prediction (GS/GP) has been widely applied in plant breeding. Arabidopsis thaliana, as a common model organism, provides important resources for dissecting genetic variation and evolutionary mechanisms of complex traits. Quantitative traits are typically influenced by multiple minor-effect genes, which are often functionally related and can be enriched within gene ontology (GO) pathways. However, optimizing marker subsets associated with these pathways to enhance GP performance remains challenging. In this study, we propose an improved GS framework called binGO-GS by integrating GO-based biological priors with a novel bin-based combinatorial SNP subset selection strategy. We evaluated the performance of binGO-GS on nine quantitative traits from two A. thaliana datasets, comprising nearly 1,000 samples and over 1.8 million SNPs. Compared with using either the full marker set or randomly selected markers with Genomic BLUP (GBLUP), binGO-GS achieved statistically significant improvements in prediction accuracy across all traits. Similar improvements were observed across six additional regression models when applying binGO-GS instead of the full marker set. Furthermore, the selected markers for identical or similar morphological traits exhibited consistent patterns in quantity and genomic distribution, supporting the polygenic model of complex quantitative traits driven by minor-effect genes. Taken together, binGO-GS offers a powerful and interpretable approach to enhance GS performance, providing a methodological reference for accelerating plant breeding and germplasm innovation.
- Open Access
- 29 October 2024
- Published by MDPI AG in Agriculture
- Vol. 14 (11) , 1915
GDnet-IP: Grouped Dropout-Based Convolutional Neural Network for Insect Pest Recognition
Lightweight convolutional neural network (CNN) models have proven effective in recognizing common pest species, yet challenges remain in enhancing their nonlinear learning capacity and reducing overfitting. This study introduces a grouped dropout strategy and modifies the CNN architecture to improve the accuracy of multi-class insect recognition. Specifically, we optimized the base model by selecting appropriate optimizers, fine-tuning the dropout probability, and adjusting the learning rate decay strategy. Additionally, we replaced ReLU with PReLU and added BatchNorm layers after each Inception layer, enhancing the model’s nonlinear expression and training stability. Leveraging the Inception module’s branching structure and the adaptive grouping properties of the WeDIV clustering algorithm, we developed two grouped dropout models, the iGDnet-IP and GDnet-IP. Experimental results on a dataset containing 20 insect species (15 pests and five beneficial insects) demonstrated an increase in cross-validation accuracy from 84.68% to 92.12%, with notable improvements in the recognition rates for difficult-to-classify species, such as Parnara guttatus Bremer and Grey (PGBG) and Papilio xuthus Linnaeus (PXLL), increasing from 38% and 47% to 62% and 93%, respectively. Furthermore, these models showed significant accuracy advantages over standard dropout methods on test sets, with faster training times compared to four conventional CNN models, highlighting their suitability for mobile applications. Theoretical analyses of model gradients and Fisher information provide further insight into the grouped dropout strategy’s role in improving CNN interpretability for insect recognition tasks.
- 29 October 2024
- Published by MDPI AG in Agriculture
- Vol. 14 (11) , 1915
Open AccessGDnet-IP: Grouped Dropout-Based Convolutional Neural Network for Insect Pest Recognition
Lightweight convolutional neural network (CNN) models have proven effective in recognizing common pest species, yet challenges remain in enhancing their nonlinear learning capacity and reducing overfitting. This study introduces a grouped dropout strategy and modifies the CNN architecture to improve the accuracy of multi-class insect recognition. Specifically, we optimized the base model by selecting appropriate optimizers, fine-tuning the dropout probability, and adjusting the learning rate decay strategy. Additionally, we replaced ReLU with PReLU and added BatchNorm layers after each Inception layer, enhancing the model’s nonlinear expression and training stability. Leveraging the Inception module’s branching structure and the adaptive grouping properties of the WeDIV clustering algorithm, we developed two grouped dropout models, the iGDnet-IP and GDnet-IP. Experimental results on a dataset containing 20 insect species (15 pests and five beneficial insects) demonstrated an increase in cross-validation accuracy from 84.68% to 92.12%, with notable improvements in the recognition rates for difficult-to-classify species, such as Parnara guttatus Bremer and Grey (PGBG) and Papilio xuthus Linnaeus (PXLL), increasing from 38% and 47% to 62% and 93%, respectively. Furthermore, these models showed significant accuracy advantages over standard dropout methods on test sets, with faster training times compared to four conventional CNN models, highlighting their suitability for mobile applications. Theoretical analyses of model gradients and Fisher information provide further insight into the grouped dropout strategy’s role in improving CNN interpretability for insect recognition tasks.
- 1 December 2021
- Published by Elsevier in Journal of Affective Disorders
- Vol. 295, 1040-1048
Improving depression prediction using a novel feature selection algorithm coupled with context-aware analysis
- 1 December 2021
- Published by Elsevier in Journal of Affective Disorders
- Vol. 295, 1040-1048
Improving depression prediction using a novel feature selection algorithm coupled with context-aware analysis