华北理工大学学报(自然科学版)

基于无人机获取的互花米草冠层高光谱影像和实测土壤有机碳(SOC)含量数据，使用数学变换和小波变换对高光谱进行变换处理，对不同尺度的小波基函数进行优选，使用竞争性自适应重加权算法(CARS)对不同变换处理后的特征光谱予以筛选，极端梯度提升(XGBoost)算法来构建土壤有机碳含量的高光谱预测模型。结果表明，小波变换最优分解尺度为coif5(L3),db4(L2),gaus4(L2),Haar(L2),mexh(L1),morl(L3),sym8(L3)。相比于数学变换，小波变换后的光谱效果预测性能更佳。其中，gaus4小波基函数构建的SOC预测模型表现出了最高的精度，测试集R²为0.479,RMSE为5.451,MAE为4.230，泛化能力相对较强。

Based on the hyperspectral image of Spartina alterniflora canopy obtained by UAV and the measured soil organic carbon(SOC) content data, the mathematical transform and wavelet transform were used to transform the hyperspectral, and the wavelet basis functions of different scales were optimized. The competitive adaptive reweighted algorithm(CARS) was used to screen the characteristic spectra after different transformation processing, and the extreme gradient boosting(XGBoost) algorithm was used to construct the hyperspectral prediction model of soil organic carbon content. The results indicate that the optimal decomposition scales of wavelet transform are coif5(L3), db4(L2), gaus4(L2), Haar(L2), mexh(L1), morl(L3), and sym8(L3). Compared with mathematical transformation, wavelet transformation demonstrates superior spectral effect prediction performance. Specifically, the SOC prediction model constructed using the gaus4 wavelet basis function achieves the highest accuracy, with the test set R² reaching 0.479, RMSE measuring 5.451, and MAE at 4.230. Additionally, the model exhibits relatively strong generalization capability.