基于热点效应融合多角度信息新指数的植被覆盖下不同深度土壤含盐量估计

Estimation of Soil Salinity at Different Depths under Vegetation Cover Using New Index Fused with Multi-angle Information Based on Hotspot Effect

摘要:

及时、准确、高效监测土壤含盐量(Soil salt content, SSC),对于实现盐渍化治理至关重要。 针对基于单一正射影像的土壤含盐量监测模型在作物覆盖条件下精度不足问题,本研究基于冠层光谱二向反射特性的热点效应可增强光谱信息理论,提出融合正射、热点和暗点 3 个观测方向的光谱数据,构建新指数提高不同土壤深度下含盐量监测精度。 采集研究区 3 块试验地空中和田间数据,构建 8 种不同类型新指数,并与 14 种常用光谱指数进行对比,通过皮尔逊相关系数法评估新建指数应用潜力。 基于特征权重算法(Resampling and feature elimination, Relief)与主成分分析(Principal component analysis, PCA) 相结合方法筛选特征变量数量,采用广义回归神经网络(General,regression neural network, GRNN)、随机森林(Random forest, RF)和极限梯度提升(Extreme gradient boosting, XG-BOOST)3 种机器学 习 算 法 构 建 4 个 不 同 土 壤 深 度 下 土 壤 含 盐 量 估 计 模 型, 通 过 决 定 系 数 ( Coefficient of determination, R2)、均方根误差(Root mean square error, RMSE)和平均绝对误差(Mean absolute error, MAE)评估模 型性能。 结果表明:对不同土壤深度,新建指数相较光谱指数与含盐量的相关性更强,最大相关性系数绝对值分别提升 44.0% 、72.9% 、70.0% 和 67.7% 。 基于 Relief 和 PCA 结合的特征筛选,可有效剔除冗余信息,减少变量数量,提高模型效率。 表层 0 ~ 10 cm 和 10 ~ 20 cm 模型精度远高于 20 ~ 40 cm 和 40 ~ 60 cm,最佳估计深度出现在 10 ~20 cm。GRNN 算法在不同土壤深度下表现均优于 RF 和 XG BOOST 算法,在最佳深度下,GRNN 算法构建模型测试集 R2为 0.802,RMSE 为 0.046% ,MAE 为 0.032% 。 研究结果为监测作物覆盖下不同深度土壤含盐量提供了新思路。

Abstract:

Timely, accurate and efficient monitoring of soil salt content ( SSC) is critical for the implementation of salinization remediation.Aiming at the insufficient accuracy of SSC estimation models established solely using nadir-view images under crop coverage conditions, relied on the theory that the hotspot effect within the bidirectional reflectance characteristics of canopy spectra can strengthen spectral information, a novel index integrating spectral data from three observation directions (nadir, hotspot and darkspot) was proposed to improve the estimation accuracy of soil salt content at various soil depths.Aerial and field datasets were collected from three experimental plots in the study area, based on which eight indices of different types were constructed and further compared with 14 commonly used spectral indices.The Pearson correlation coefficient was adopted to evaluate the application potential of the newly developed indices.Subsequently, a hybrid feature screening method combining the resampling and feature elimination (Relief) algorithm and principal component analysis (PCA) was used to filter feature variables.Three machine learning algorithms, namely general regression neural network ( GRNN), random forest (RF) and extreme gradient boosting (XG BOOST), were employed to establish SSC estimation models corresponding to four distinct soil layers.Finally, the coefficient of determination(R2)), root mean square error ( RMSE) and mean absolute error ( MAE) were utilized to quantify model performance.The results revealed that the newly constructed indices exhibited stronger correlations with soil salt content than conventional spectral indices across all soil depths.The maximum absolute values of correlation coefficients were increased by 44.0% , 72.9% , 70.0% and 67.7% , respectively.The Relief PCA combined feature selection strategy effectively eliminated redundant information, reduced the number of input variables and accelerated model operation efficiency.Model accuracy for the topsoil layers of 0 ~ 10 cm and 10 ~ 20 cm was markedly higher than that for the 20 ~ 40 cm and 40 ~ 60 cm layers, with the optimal estimation depth identified as 10 ~ 20 cm.GRNN outperformed RF and XG BOOST at all measured soil depths.At the optimal depth, the GRNN model achieved a test set R2 of 0.802, an RMSE of 0.046% and an MAE of 0.032% .The research result can provide a novel strategy for estimating soil salt content at multiple depths under crop canopy coverage.