Scholarlink

In this paper, we propose an enhanced deep learning approach for early-stage rice yield prediction by embedding self-attention layers into established deep neural networks (DNNs) such as VGG-16, DenseNet, MobileNet, ResNet, Inception, and Xception. The addition of self-attention significantly improves the models’ ability to capture long-range dependencies and global context, which traditional convolutional layers often fail to represent adequately due to their inherently local receptive fields. We evaluate the proposed method on a dataset of 18,642 RGB images collected from 47 rice field plots spanning over 28 hectares in An Giang and Tra Vinh provinces. The images were acquired using digital cameras, smartphones, and fixed-wing UAVs during the heading stage of rice growth. Experimental results demonstrate that DNNs enhanced with self-attention layers consistently outperform their original fine-tuned counterparts. Furthermore, these hybrid models also achieve higher prediction accuracy compared to Vision Transformers (ViT), highlighting the effectiveness of integrating self-attention into conventional DNN architectures for agricultural yield forecasting.

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

Embedding Self-attention Blocks into Deep Neural Networks for Early Rice Yield Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Thuy-Vi Thi Ha,
  • Phuoc-Hung Vo,
  • Thanh-Nghi Do

摘要

In this paper, we propose an enhanced deep learning approach for early-stage rice yield prediction by embedding self-attention layers into established deep neural networks (DNNs) such as VGG-16, DenseNet, MobileNet, ResNet, Inception, and Xception. The addition of self-attention significantly improves the models’ ability to capture long-range dependencies and global context, which traditional convolutional layers often fail to represent adequately due to their inherently local receptive fields. We evaluate the proposed method on a dataset of 18,642 RGB images collected from 47 rice field plots spanning over 28 hectares in An Giang and Tra Vinh provinces. The images were acquired using digital cameras, smartphones, and fixed-wing UAVs during the heading stage of rice growth. Experimental results demonstrate that DNNs enhanced with self-attention layers consistently outperform their original fine-tuned counterparts. Furthermore, these hybrid models also achieve higher prediction accuracy compared to Vision Transformers (ViT), highlighting the effectiveness of integrating self-attention into conventional DNN architectures for agricultural yield forecasting.