Scholarlink

<p>Agriculture is vital for global food security and economic stability. Technological advancements can significantly enhance productivity and sustainability. Mango, one of the world’s most cherished fruits, holds immense commercial and nutritional value. However, its cultivation is threatened by diseases such as Alternaria (causing dark leaf spots), Anthracnose (leading to black, sunken spots), Black Mould Rot (characterized by sooty fungal growth), Stem End Rot (causing decay from the stem end), and the crucial Healthy category. These diseases compromise fruit quality and yield, causing substantial economic losses. Manual diagnosis is subjective, time-consuming, and requires expertise, making it impractical for large-scale use. Artificial intelligence offers a&#xa0;transformative solution by enabling automated, accurate disease identification. In deep learning, conventional models such as Convolutional Neural Networks (CNNs) have a&#xa0;limited receptive field and struggle with long-range dependencies. Also, vision transformers are data-hungry and computationally exhaustive. Both of the above approaches lack focus on disease-specific features in complex agricultural settings. To address these gaps, we propose SAAEffNet, a&#xa0;synthetic-augmented attention-based EfficientNet deep learning model for classifying mango diseases. The model begins with a&#xa0;curated dataset, expanded using CycleGAN to generate high-fidelity synthetic images for enhanced diversity. In the previous pre-trained model, i.e., EfficientNetB3, we integrate the Convolutional Block Attention Module (CBAM) approach to refine features such as dark leaf spots, sunken spots, sooty fungal growth, etc. The model is trained with adaptive scheduling and early stopping. Our model achieves superior generalization, offering a&#xa0;robust tool for automated mango disease diagnosis. The model has undergone rigorous training and validation with a&#xa0;comprehensive dataset that includes images of both diseased and healthy mangoes, ensuring its effectiveness in accurately detecting and diagnosing mango health conditions. The results clearly illustrate our model’s exceptional performance, achieving remarkable metrics: accuracy 98.40%, precision 98.40%, recall 98.40%, and F1-score 98.39%, all with a&#xa0;minimal loss of just 1%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SAAEffNet: A Synthetic-Augmented Attention EfficientNet Model for Mango Disease Classification

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Anil Sandhi,
  • Rajeev Kumar,
  • Reeta Bhardwaj,
  • Dinesh Kumar

摘要

Agriculture is vital for global food security and economic stability. Technological advancements can significantly enhance productivity and sustainability. Mango, one of the world’s most cherished fruits, holds immense commercial and nutritional value. However, its cultivation is threatened by diseases such as Alternaria (causing dark leaf spots), Anthracnose (leading to black, sunken spots), Black Mould Rot (characterized by sooty fungal growth), Stem End Rot (causing decay from the stem end), and the crucial Healthy category. These diseases compromise fruit quality and yield, causing substantial economic losses. Manual diagnosis is subjective, time-consuming, and requires expertise, making it impractical for large-scale use. Artificial intelligence offers a transformative solution by enabling automated, accurate disease identification. In deep learning, conventional models such as Convolutional Neural Networks (CNNs) have a limited receptive field and struggle with long-range dependencies. Also, vision transformers are data-hungry and computationally exhaustive. Both of the above approaches lack focus on disease-specific features in complex agricultural settings. To address these gaps, we propose SAAEffNet, a synthetic-augmented attention-based EfficientNet deep learning model for classifying mango diseases. The model begins with a curated dataset, expanded using CycleGAN to generate high-fidelity synthetic images for enhanced diversity. In the previous pre-trained model, i.e., EfficientNetB3, we integrate the Convolutional Block Attention Module (CBAM) approach to refine features such as dark leaf spots, sunken spots, sooty fungal growth, etc. The model is trained with adaptive scheduling and early stopping. Our model achieves superior generalization, offering a robust tool for automated mango disease diagnosis. The model has undergone rigorous training and validation with a comprehensive dataset that includes images of both diseased and healthy mangoes, ensuring its effectiveness in accurately detecting and diagnosing mango health conditions. The results clearly illustrate our model’s exceptional performance, achieving remarkable metrics: accuracy 98.40%, precision 98.40%, recall 98.40%, and F1-score 98.39%, all with a minimal loss of just 1%.