Scholarlink

Accurate rice yield prediction before harvest is crucial for optimizing crop management, ensuring food security, and informing trade and policy decisions. In this paper, we propose a novel approach that combines Vision Transformer (ViT) feature extraction with a newly designed local learning algorithm to enhance rice yield prediction from field images taken at the ripening stage. Unlike conventional deep learning models, our method replaces the standard ViT classification head (Multi-Layer Perceptron - MLP) with a custom regression layer to predict numeric yield values. Furthermore, we introduce a local learning algorithm that finds k nearest neighbors from the training dataset for each test instance and trains a local regression model to improve prediction accuracy. Experimental results on image datasets collected from 47 plots covering over 28 hectares in the Mekong Delta region demonstrate the effectiveness of our approach. The proposed local learning model, leveraging ViT-extracted features, achieved the lowest mean absolute error (MAE) of 72.49 kgs/1000 \(m^2\) , outperforming conventional regression methods. These findings confirm the viability of integrating ViT-based feature extraction with local learning techniques for precise and adaptive rice yield prediction.

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

Integrating Local Learning Algorithms and Vision Transformers for Accurate Rice Yield Prediction

  • 会议论文

  • 发表日期：

    2026年

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

摘要

Accurate rice yield prediction before harvest is crucial for optimizing crop management, ensuring food security, and informing trade and policy decisions. In this paper, we propose a novel approach that combines Vision Transformer (ViT) feature extraction with a newly designed local learning algorithm to enhance rice yield prediction from field images taken at the ripening stage. Unlike conventional deep learning models, our method replaces the standard ViT classification head (Multi-Layer Perceptron - MLP) with a custom regression layer to predict numeric yield values. Furthermore, we introduce a local learning algorithm that finds k nearest neighbors from the training dataset for each test instance and trains a local regression model to improve prediction accuracy. Experimental results on image datasets collected from 47 plots covering over 28 hectares in the Mekong Delta region demonstrate the effectiveness of our approach. The proposed local learning model, leveraging ViT-extracted features, achieved the lowest mean absolute error (MAE) of 72.49 kgs/1000 \(m^2\) , outperforming conventional regression methods. These findings confirm the viability of integrating ViT-based feature extraction with local learning techniques for precise and adaptive rice yield prediction.