| 基于ResNet-MHAM模型的山区耕地土壤有机质含量高光谱反演 |
| 摘要点击 478 全文点击 58 投稿时间：2024-02-29 修订日期：2024-06-02 |
| 查看HTML全文
查看全文 查看/发表评论 下载PDF阅读器 |
| 中文关键词 高光谱 残差网络（ResNet） 多头注意力机制（MHAM） 土壤有机质（SOM） 山区耕地 |
| 英文关键词 hyperspectral residual network （ResNet） multi-head attention mechanism （MHAM） soil organic matter （SOM） mountainous farmland |
|
| 中文摘要 |
| 针对贵州喀斯特山区耕地土壤有机质（SOM）含量高光谱遥感预测的精度和泛化能力不足的问题,提出了结合残差网络（ResNet）和多头注意力机制（MHAM）的一维高光谱反射数据模型（ResNet-MHAM）. 首先,采集贵州13个县市区188个土壤样品并检测光谱信息；其次,基于不同层数（34、50、101和152层）的ResNet结构并结合MHAM进行优化构建模型；最后,使用30%的数据集和十折交叉验证进行模型验证. 实验结果显示,50层ResNet结构与MHAM的结合模型,在决定系数（R2）达到0.917 2,均方根误差（RMSE）为7.454 9 g·kg-1,表现出优于BPNN、SVM、PLSR、GPR和RF模型的准确性和泛化能力. 研究结果为贵州山区SOM含量的高光谱预测提供了新的有效方法. |
| 英文摘要 |
| In response to the lack of accuracy and generalization challenges in predicting soil organic matter （SOM） content in the karst mountainous agricultural soils of the Guizhou Province using hyperspectral remote sensing, a one-dimensional hyperspectral reflectance data model, termed ResNet-MHAM, was proposed. First, soil samples from 188 locations across 13 counties and districts in Guizhou were collected, and their spectral information was analyzed. Second, the ResNet structure was optimized in combination with MHAM across different layers （34, 50, 101, and 152 layers） to construct the model presented in this study. Finally, model validation was conducted using 30% of the dataset and 10-fold cross-validation. Experimental results demonstrated that the optimized version of the model combining 50-layer ResNet structure with MHAM achieved a coefficient of determination （R2） of 0.917 2 and a root mean square error （RMSE） of 7.454 9 g·kg-1, showcasing superior accuracy and generalization capabilities compared to commonly used models such as BPNN, SVM, PLSR, GPR, and RF. These findings provide a novel and effective approach for hyperspectral prediction of SOM content in the mountainous regions of Guizhou. |
|
|
|
