Scholarlink

<p>This study aims to improve the estimation of soil organic carbon (SOC) content by integrating visible-near-infrared (Vis-NIR) spectral information and sample location information through a multi-modal deep learning framework, termed SpatialFormer. The proposed two-stage calibration model first transforms one-dimensional Vis-NIR spectral data into two-dimensional spectrograms, which are then modelled using a Vision Transformer (ViT) to capture complex spectral features. In the second stage, the residuals from the spectral predictions are modelled through spatially dependent neural networks, allowing for improved final estimation. The model was validated using the LUCAS dataset at both the European and French scales, achieving an R-squared (R<sup>2</sup>) of 0.78, root mean square error (RMSE) of 9.09&#xa0;g·kg⁻¹, and relative percent difference (RPD) of 2.12 at the European scale, and an R² of 0.71, RMSE of 8.33&#xa0;g·kg⁻¹, and RPD of 1.87 at the national scale. Comparative analyses reveal that the proposed model outperforms those relying solely on spectral data, demonstrating its potential for improving SOC estimation and supporting precision agriculture and environmental management.</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

SpatialFormer: A Model to Estimate Soil Organic Carbon Content Using Spectral and Spatial Information

  • 期刊论文

  • 发表日期：

    2025年2月25日

查看全文
  • Zichen Tong,
  • Lanfa Liu

摘要

This study aims to improve the estimation of soil organic carbon (SOC) content by integrating visible-near-infrared (Vis-NIR) spectral information and sample location information through a multi-modal deep learning framework, termed SpatialFormer. The proposed two-stage calibration model first transforms one-dimensional Vis-NIR spectral data into two-dimensional spectrograms, which are then modelled using a Vision Transformer (ViT) to capture complex spectral features. In the second stage, the residuals from the spectral predictions are modelled through spatially dependent neural networks, allowing for improved final estimation. The model was validated using the LUCAS dataset at both the European and French scales, achieving an R-squared (R2) of 0.78, root mean square error (RMSE) of 9.09 g·kg⁻¹, and relative percent difference (RPD) of 2.12 at the European scale, and an R² of 0.71, RMSE of 8.33 g·kg⁻¹, and RPD of 1.87 at the national scale. Comparative analyses reveal that the proposed model outperforms those relying solely on spectral data, demonstrating its potential for improving SOC estimation and supporting precision agriculture and environmental management.