Abstract Rapid analytical methods are needed to measure organic carbon (OC) and total nitrogen (Nt) contents in reclaimed mine soils. Near infrared spectroscopy (NIRS) with appropriate chemometric techniques could be used for OC and Nt monitoring in the mine soils. The aim of this study was to compare efficiency of NIR-based models developed using various chemometric approaches to predict OC and Nt contents in afforested mine soils. The studied approaches were: partial least square regression (PLSR), principal component regression (PCR) and artificial neural networks based on the entire spectral data (ANN), the principal components (PCA-ANN) and the latent variables (PLS-ANN) calculated from the spectral data. The samples (n = 90) of uppermost mine soil horizons (0–20 cm) were taken from the reclaimed dump of Belchatow lignite mine (Poland) and measured for the OC content by dry combustion and for the Nt content by Kjehldahl method. The samples were air-dried, finely ground and their NIR spectra (1000 nm–2500 nm) were recorded. The models were developed using 60 samples while the remaining 30 samples were used for independent validation. All the tested chemometric approaches (except the ANN model for OC) yielded excellent models useful for quantitative estimations. The PLSR models were promising at the calibration stage (values of ratio of inter-quartile distance to standard error of prediction (RPIQC) were 7.00 and 4.22 for Nt and OC, respectively) however in the validation they performed less successfully (RPIQV = 3.08 for Nt and RPIQV = 2.75). The accuracy of ANN models based on the entire spectra was similar to PLSR or PCR models. However, the ANN based on reduced spectral data (PCA-ANN and PLS-ANN) performed distinctly better. The most accurate predictive models for the OC and Nt contents were obtained using PCA-ANN approach (RPIQV = 3.64 and 2.90, for Nt and OC, respectively). The results indicate that NIRS coupled with ANN based on the reduced spectral data can be successfully applied to measure the OC and Nt contents in mine soils.

中文翻译：

---

用近红外反射光谱和各种化学计量学方法估算矿井土壤中的总氮和有机碳含量

摘要 需要快速分析方法来测量再生矿土中的有机碳 (OC) 和总氮 (Nt) 含量。具有适当化学计量学技术的近红外光谱 (NIRS) 可用于矿山土壤中的 OC 和 Nt 监测。本研究的目的是比较使用各种化学计量学方法开发的基于 NIR 的模型的效率，以预测绿化矿山土壤中的 OC 和 Nt 含量。研究的方法是：偏最小二乘回归 (PLSR)、主成分回归 (PCR) 和基于整个光谱数据 (ANN)、主成分 (PCA-ANN) 和潜在变量 (PLS-ANN) 的人工神经网络从光谱数据计算。从 Belchatow 褐煤矿（波兰）的再生堆中取出矿井最上层土壤层（0-20 厘米）的样品（n = 90），并通过干燃烧法测量 OC 含量，并通过凯氏法测量 Nt 含量。将样品风干、精细研磨并记录其 NIR 光谱（1000 nm–2500 nm）。这些模型是使用 60 个样本开发的，而其余 30 个样本用于独立验证。所有经过测试的化学计量方法（除了用于 OC 的 ANN 模型）都产生了对定量估计有用的优秀模型。PLSR 模型在校准阶段很有希望（Nt 和 OC 的四分位距与标准预测误差的比值 (RPIQC) 的值分别为 7.00 和 4.22），但在验证中它们的表现不太成功（RPIQV = 3.08，对于 Nt 和 OC） Nt 和 RPIQV = 2.75)。基于整个光谱的 ANN 模型的准确性类似于 PLSR 或 PCR 模型。然而，基于简化频谱数据的 ANN（PCA-ANN 和 PLS-ANN）表现明显更好。OC 和 Nt 含量的最准确预测模型是使用 PCA-ANN 方法获得的（Nt 和 OC 分别为 RPIQV = 3.64 和 2.90）。结果表明，基于简化光谱数据的 NIRS 结合 ANN 可以成功地应用于测量矿井土壤中的 OC 和 Nt 含量。