基于遥感时-空-谱特征及随机森林模型的土壤重金属空间分布预测

获取土壤重金属的含量特征及空间分布是预防土壤污染和制定环保政策的关键.选取济南市长清区为研究区，系统采集304处表层土壤样品（0~20 cm），利用多源遥感数据构建土壤重金属的光谱特征、时间特征和空间特征；进一步采用相关分析法选择出与土壤重金属密切相关的时-空-谱特征，并将其作为输入自变量，实测土壤砷（As）含量值为因变量，建立基于随机森林（RF）算法的空间预测模型，完成土壤重金属的含量估算和空间分布预测.结果表明：①As含量均值超出背景值43.17％，低于农用地土壤污染风险规定的筛选值和管控值，表明As在土壤中出现富集，但处于可管控范围内.②在单个遥感特征构建的土壤重金属空间预测模型中，精度由高到低依次为：空间特征（RPIQ=3.87）>时间特征（RPIQ=2.57）>光谱特征（RPIQ=2.50），空间特征对土壤重金属空间预测最为重要.③基于“时间-空间”、“时间-光谱”和“空间-光谱”组合特征的土壤重金属空间预测模型均优于单个特征构建的模型，其精度系数RPIQ值分别为4.81、4.21和4.70.④利用“时间-空间-光谱”特征组合输入的随机森林模型达到最佳的空间预测精度（R²=0.90；RMSE=0.77；RPIQ=5.68）.⑤As在空间分布上从西北到东南含量逐步降低，主要受到黄河冲淤积和工业活动影响.研究采用的遥感时-空-谱特征结合随机森林算法的土壤重金属空间预测技术，可为土壤污染防治及环境风险管控提供有效的方法支持.

Distribution Prediction of Soil Heavy Metals Based on Remote Sensing Temporal-Spatial-Spectral Features and Random Forest Model

Obtaining soil heavy metal content characteristics and spatial distribution is crucial for preventing soil pollution and formulating environmental protection policies. We collected 304 surface soil samples (0-20 cm) in the Changqing district. At the same time, the spectral, temporal, and spatial features of soil heavy metals were derived from multi-remote sensing data; the temporal-spatial-spectral features closely related to soil heavy metals were selected via correlation analysis and used as input independent variables. The measured soil arsenic (As) content was used as the dependent variable to establish a spatial prediction model based on the random forest (RF) algorithm. The results showed the following:the As content in the soils exceeded the background value by 43.17% but did not exceed the risk screening values and intervention values, indicating slight heavy metal pollution in the soil. The accuracy ranking of the spatial prediction models with one feature type from high to low was spatial features (ratio of performance to inter-quartile range (RPIQ)=3.87)>temporal features (RPIQ=2.57)>spectral features (RPIQ=2.50). The spatial features were the most informative for predicting soil heavy metals. The models using temporal-spatial, temporal-spectral, and spatial-spectral features were superior to those using only one feature type, and the RPIQ values were 4.81, 4.21, and 4.70, respectively. The RF model with temporal-spatial-spectral features achieved the highest spatial prediction accuracy (R²=0.90; root mean square error (RMSE)=0.77; RPIQ=5.68). The As content decreased from the northwest to the southeast due to Yellow River erosion and industrial activities. The spatial prediction of soil heavy metals incorporating remote sensing temporal-spatial-spectral features and the random forest model provides effective support for soil pollution prevention and environmental risk control.