针对土壤高光谱数据量大、存在光谱信息冗余和重叠现象, 应用稳定竞争性自适应重加权采样策略挑选特征变量, 结合偏最小二乘回归和随机森林建立土壤有机质含量估算模型, 并与竞争性自适应重加权算法、迭代保留有效信息变量、连续投影算法和遗传算法所得结果进行比较。结果显示, 5种变量选择算法挑选的特征变量主要分布在1 900～2 400 nm的近红外光谱区域。RF模型的预测效果优于PLSR模型; 与PLSR模型相比, RF模型鲁棒性更好, 对异常值和噪声的敏感度更低。基于sCARS算法挑选的特征变量建立RF模型, 变量数为51个, 仅占全波段的2.55%, 验证集R2=0.958, 获得的RPD为4.7, 能够很好地预测SOM含量。

Aiming at the phenomenon of large amount of soil hyperspectral data, spectral information redundancy and overlap, a model for estimating soil organic matter content is established by using a stable competitive adaptive reweighting sampling strategy and combining partial least squares regression with random forest. The results are compared with those obtained by competitive adaptive reweighting algorithm, iterative iteratively retains information variables, successive projections algorithm and genetic algorithm. The five variable selection algorithms selecting the characteristic variables mainly distribute in the near infrared spectral region of 1 900-2 400 nm. The prediction effect of the RF model is better than that of the PLSR model. Compared with the PLSR model, the RF model has better robustness, lower sensitivity to outliers and noise. The RF model is established based on the characteristic variables selected by sCARS algorithm. The number of variables is 51, accounting for only 2.55% of the whole band. The verification set R2 is 0.958, and the RPD is 4.7, which can predict the SOM content very well.