【摘要】 土壤有机质（SOM）含量是衡量土壤肥力的重要指标之一，从高光谱遥感图像中有效预测SOM含量具有重要意义。传统的机器学习方法需要复杂的特征工程且精度不高，而以卷积神经网络（CNN）为代表的深度学习方法在土壤高光谱领域研究较少，且对小样本数据建模精度较差，光谱数据的空间特征提取不足。因此，提出了一种使用通道注意力机制的一维空洞卷积网络模型（SE-DCNN）。以广西国有黄冕林场和国有雅长林场采集的207个土壤样本为研究对象，对比分析了3种机器学习方法和4种深度学习方法在不同光谱预处理下的建模效果。结果表明，SE-DCNN模型因为使用了空洞卷积和通道注意力机制，扩大感受野并提取多尺度特征，有较好的建模精确度和泛化拟合能力。最佳预测模型是基于S-G降噪（SGD）和一阶微分（DR）的光谱预处理方式建立的SE-DCNN模型，验证集的决定系数（R²）为0.971,均方根误差（RMSE）为2.042 g·kg^-1,相对分析误差（RPD）为5.273。因此，使用SE-DCNN能够对广西林地红壤有机质含量进行准确预测。更多还原

【Abstract】 Soil Organic Matter（SOM） content is one of the important indicators used to measure soil fertility, and it is of great significance in accurately predicting SOM content from hyperspectral remote sensing images. Traditional machine learning methods require complex feature engineering. Still, they are not highly accurate, while deep learning methods represented by Convolutional Neural Networks（CNNs） are less studied in soil hyperspectral, and the modeling accuracy of small sample data is poor. The spatial feature extraction of spectral data is insufficient. This paper proposes a one-dimensional convolutional network model using a channel attention mechanism（SE Dilated Convolutional Neural Network, SE-DCNN）. Taking 207 soil samples collected from Guangxi State-owned Huangmian Forest Farm and State-owned Yachang Forest Farm as research objects, this paper compares and analyzes the modeling effects of 3 machine learning and 4 deep learning methods under different spectral preprocessing. The results show that the SE-DCNN model, because of the use of dilated convolution and channel attention mechanism, expands the receptive field, extracts multi-scale features, and has good modeling accuracy and generalization fitting ability. The best prediction model in this paper is the SE-DCNN model established based on the spectral preprocessing method of Savitaky-Golay denoising（SGD） and first-order derivative（DR）, the determination coefficient（R²） of the validation set is 0.971, the root mean square error（RMSE） is 2.042 g·kg^-1, and the relative analysis error（RPD） is 5.273. Therefore, SE-DCNN can accurately predict the organic matter content of red soil in Guangxi forest land.更多还原