Scholarlink

<p>Agriculture, land management, and environmental monitoring depend on precise soil categorization for crop selection, irrigation, and sustainable land use. This study introduces a hybrid soil classification system that uses lightweight deep learning models, attention mechanisms, and manually created colour features for resource-limited contexts. A customized soil image dataset with five classes—Black, Cinder, Laterite, Peat, and Yellow—was used for training and assessment. A modified MobileNetv2 was used in the initial stage for its efficiency and suitability for edge devices. In soil images, this model identified texture and spatial patterns with 84% accuracy. However, certain misclassifications of similar soil types highlighted the need for better feature representation. Adding deep MobileNetv2 features and colour features improved the technique, and a multi-class support vector machine (SVM) with Error-Correcting Output Codes classified the data. This fusion strategy improved classification performance by 96%, showing the benefits of using both deep and handcrafted features with a robust classifier like SVM. The methodology can be broadened by integrating bigger and more diverse datasets along with geolocation information to enhance scalability and adaptability.</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

A hybrid approach for soil classification: integrating lightweight convolutional neural network, attention mechanisms, and color features under resource constraints

  • 期刊论文

  • 发表日期：

    2025年12月20日

查看全文
  • Santi Kumari Behera,
  • Sachin Vishwakarma,
  • V. V. Teresa,
  • Ashoka Kumar Ratha,
  • Prabira Kumar Sethy,
  • Aziz Nanthaamornphong

摘要

Agriculture, land management, and environmental monitoring depend on precise soil categorization for crop selection, irrigation, and sustainable land use. This study introduces a hybrid soil classification system that uses lightweight deep learning models, attention mechanisms, and manually created colour features for resource-limited contexts. A customized soil image dataset with five classes—Black, Cinder, Laterite, Peat, and Yellow—was used for training and assessment. A modified MobileNetv2 was used in the initial stage for its efficiency and suitability for edge devices. In soil images, this model identified texture and spatial patterns with 84% accuracy. However, certain misclassifications of similar soil types highlighted the need for better feature representation. Adding deep MobileNetv2 features and colour features improved the technique, and a multi-class support vector machine (SVM) with Error-Correcting Output Codes classified the data. This fusion strategy improved classification performance by 96%, showing the benefits of using both deep and handcrafted features with a robust classifier like SVM. The methodology can be broadened by integrating bigger and more diverse datasets along with geolocation information to enhance scalability and adaptability.