Estimation of SOC using VNIR and MIR hyperspectral data based on spectral-to-image transforming and multi-channel CNN

文献类型： 外文期刊

第一作者： Tang, Aohua

作者： Tang, Aohua;Yang, Guijun;Li, Zhenhong;Chen, Weinan;Zhang, Jing;Tang, Aohua;Yang, Guijun;Pan, Yuchun;Liu, Yu;Long, Huiling;Chen, Weinan;Zhang, Jing;Yang, Yue;Yang, Xiaodong;Xu, Bo;Yang, Yue

作者机构：

关键词： MIR spectral; Multi-channel-CNN; SIT; Soil organic carbon; VNIR spectral

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 231 卷

页码：

收录情况： SCI

摘要： Soil organic carbon (SOC) plays a crucial role in soil functions, ecosystem health, and carbon cycling. Accurately estimating SOC content is essential for sustainable agricultural production. Traditional spectral methods, particularly visible-near-infrared (VNIR) and mid-infrared (MIR) spectroscopy, are extensively used for SOC estimation. While multi-source data fusion can improve model performance, existing fusion algorithms often struggle to effectively capture multi-band integrated features from multiple sensors. To overcome this limitation and improve the accuracy of SOC estimation, we propose a novel deep learning model, SOCNet, which integrates multi-source VNIR and MIR spectral data through the Spectral-to-Image Transform (SIT) technique. The SIT converts one-dimensional spectral into informative dual-bands images, which are subsequently processed and fused using a multi-channel convolutional neural network (MC-CNN) to enhance the precision of SOC content prediction. To assess SOCNet's performance, we compared it against three CNN models using individual SIT images (SIT-CNNs), three partial least squares regression (PLSR) models (MIR-PLSR, VNIR-PLSR, VNIR-MIRPLSR), and three 1D-CNN models using VNIR and MIR spectral data. Results showed that the SIT technique significantly enhanced the spectral responsiveness to SOC, achieving a maximum absolute Pearson correlation coefficient (|r|) of 0.96. SIT-CNN models outperformed traditional PLSR models, particularly the SIT-CNN model utilizing VNIR spectra alone, where the R2 increased from 0.79 to 0.95 and RMSE decreased from 91.52 to 43.36 g/kg. Furthermore, feature fusion at various levels within the CNN architecture led to further enhancements in SOC prediction accuracy. Among the evaluated configurations, SOCNet's Scheme II achieved the highest prediction accuracy, with an R2 of 0.99, an RMSE of 19.83 g/kg, and a MAE of 12.30 g/kg. Compared to other models, SOCNet reduced RMSE and MAE by at least 36.22 % and 38.84 %, respectively. These findings highlight that SOCNet provides a superior advantage in predicting SOC content, making it promising for applications in precision agriculture.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序