Soil hyperspectral and soil organic carbon

This study hypothesized that integrating multiscale wavelet decomposition (CWT/DWT) with machine learning (ML) enhances soil organic carbon (SOC) estimation in arid lakeside oases. Using 82 soil samples with VNIR spectra, CWT at scales 1-5 reduced noise by 19.21% vs DWT. CWT-1-CARS-RF achieved optimal accuracy (R²=0.79, RPD=2.23), with 49.04% R² and 58.23% RPD improvements via feature selection. Sensitive bands were 401-504 nm (visible) and 1638-2369 nm (NIR). Spatial validation showed 91.3% consistency, confirming robust SOC mapping via wavelet-ML synergy.

本研究提出假设：将多尺度小波分解 (Multiscale Wavelet Decomposition) 与机器学习 (ML, Machine Learning) 相结合，可提升干旱湖滨绿洲区域的土壤有机碳 (SOC, Soil Organic Carbon) 估算精度。本研究共采集82份土壤样品并获取其可见近红外光谱 (VNIR, Visible and Near-Infrared Spectroscopy) 数据；相较于离散小波变换 (DWT, Discrete Wavelet Transform)，尺度1至5的连续小波变换 (CWT, Continuous Wavelet Transform) 可降低19.21%的噪声。其中，CWT-1-CARS-RF模型取得了最优预测精度（决定系数R²=0.79，残差预测偏差RPD=2.23），通过特征选择使模型的R²与RPD分别提升49.04%与58.23%。研究识别得到的敏感波段为401~504 nm（可见光波段）与1638~2369 nm（近红外波段，NIR, Near-Infrared）。空间验证结果显示模型一致性达91.3%，证实了基于小波变换-机器学习协同方法可实现稳健的土壤有机碳制图。