Scholarlink

<p>Soil classification is a critical task in environmental science, agriculture, and land management, as it influences decision-making processes related to land use and resource management. This study presents a novel Hybrid Transfer Learning approach that integrates raster and text data collected from World Soil Database(WSB) to enhance soil classification accuracy. A comprehensive comparison of various transfer learning architectures has been conducted, including ResNet18, ResNet50, DenseNet121, VGG16, and VGG19, to evaluate their performance in soil classification tasks. Additionally, systematically analyze different combinations of activation functions (ReLU, Leaky ReLU, Sigmoid, and Tanh) and optimizers (Adam, AdaMax, RMSProp, and SGD) to identify the best configurations for high classification accuracy. To further interpret the model’s predictions, SHAP (SHapley Additive exPlanations) analysis is applied, providing insights into feature importance and model behavior. The results demonstrate that the proposed HybridTransferNet approach significantly outperforms traditional methods, achieving an impressive accuracy of 99.47%. This research not only highlights the effectiveness of hybrid models in soil classification but also offers valuable insights into the optimal configurations for leveraging transfer learning in environmental applications.</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

Soil classification revolutionized: A hybrid transfer learning approach with SHAP analysis

  • 期刊论文

  • 发表日期：

    2025年5月20日

查看全文
  • Chetan Raju,
  • D V Ashoka,
  • Ajay Prakash B V

摘要

Soil classification is a critical task in environmental science, agriculture, and land management, as it influences decision-making processes related to land use and resource management. This study presents a novel Hybrid Transfer Learning approach that integrates raster and text data collected from World Soil Database(WSB) to enhance soil classification accuracy. A comprehensive comparison of various transfer learning architectures has been conducted, including ResNet18, ResNet50, DenseNet121, VGG16, and VGG19, to evaluate their performance in soil classification tasks. Additionally, systematically analyze different combinations of activation functions (ReLU, Leaky ReLU, Sigmoid, and Tanh) and optimizers (Adam, AdaMax, RMSProp, and SGD) to identify the best configurations for high classification accuracy. To further interpret the model’s predictions, SHAP (SHapley Additive exPlanations) analysis is applied, providing insights into feature importance and model behavior. The results demonstrate that the proposed HybridTransferNet approach significantly outperforms traditional methods, achieving an impressive accuracy of 99.47%. This research not only highlights the effectiveness of hybrid models in soil classification but also offers valuable insights into the optimal configurations for leveraging transfer learning in environmental applications.