Detecting soil arsenic (As) contamination is crucial for designing efficient soil remediation strategies; however, traditional laboratory-based As detection techniques are time- and labour-intensive and are unsuitable for large-scale spatial analyses. To address this issue, we combined machine learning (ML) with visible-near-infrared (vis-NIR) spectroscopy to develop an efficient framework for As detection in soil. The optimal spectral preprocessing method was determined, and eight ML models were compared. The support vector classifier achieved optimal performance after subsequent hyperparameter tuning, with area under the curve (AUC) and accuracy values of 0.89 and 0.83, respectively. Important spectral bands at 471 and 2422 nm were identified by permutation importance and correspond to Fe-oxide and carbonate, respectively. These two wavelengths were included in the partial dependence plot (PDP), revealing that the likelihood of soil As contamination decreased with increasing reflectance at wavelengths of 471 and 2422 nm due to a decrease in Fe-oxide and carbonate content. Consistent with this finding, two-way PDP analysis revealed that the As content of soil increased with increasing Fe-oxide and carbonate content. The model’s classification performance was further improved using an ensemble technique based on three optimal ML models, resulting in increased AUC and accuracy values of 0.9 and 0.83, respectively. Overall, the framework presented in this study enabled the precise classification of soil As content at the continental scale, while also indirectly explained the complex relationships between As content and soil properties.

中文翻译：

---

通过可见近红外光谱和机器学习对欧盟土壤中的砷污染进行分类


检测土壤砷（As）污染对于设计有效的土壤修复策略至关重要；然而，传统的基于实验室的砷检测技术耗时且费力，并且不适合大规模空间分析。为了解决这个问题，我们将机器学习 (ML) 与可见光-近红外 (vis-NIR) 光谱相结合，开发了土壤中砷检测的有效框架。确定了最佳的光谱预处理方法，并对八种ML模型进行了比较。支持向量分类器经过后续的超参数调整后达到了最佳性能，曲线下面积（AUC）和准确率值分别为0.89和0.83。 471 和 2422 nm 处的重要光谱带通过排列重要性进行识别，分别对应于铁氧化物和碳酸盐。这两个波长包含在部分相关图 (PDP) 中，表明由于铁氧化物和碳酸盐含量的减少，土壤 As 污染的可能性随着 471 和 2422 nm 波长处反射率的增加而降低。与这一发现一致，双向 PDP 分析表明，土壤中砷含量随着氧化铁和碳酸盐含量的增加而增加。使用基于三个最佳 ML 模型的集成技术进一步提高了模型的分类性能，使 AUC 和准确度值分别增加了 0.9 和 0.83。总体而言，本研究提出的框架能够在大陆尺度上对土壤砷含量进行精确分类，同时也间接解释了砷含量与土壤性质之间的复杂关系。