可见-近红外光谱已被证明是一种快速、 有效的有机碳(TOC)含量预测方法。 但是, 当前利用光谱预测TOC含量的研究对象主要为土壤或湖泊沉积物, 还未见潮间带海洋沉积物的研究报道。 为了快速准确预测潮间带沉积物TOC含量, 通过异常样本剔除、 光谱特征变换、 特征波长提取相结合, 构建TOC预测模型, 即, 采集潮间带沉积物样品光谱, 采用马氏距离、 标准杠杆值和学生残差联合分析的方法剔除异常样本, 利用多元散射校正(MSC)、 平滑+微分进行光谱变换, 利用遗传算法(GA)提取特征波长, 采用偏最小二乘法(PLS)、 最小二乘支持向量机(LSSVM)和BP神经网络(BPNN)对沉积物TOC含量进行建模和预测, 通过决定系数( R²)和剩余估计偏差(PRD)来评价模型精度。 结果表明, 剔除异常样本有助于提升模型精度, BPNN模型的检验 R²和PRD分别提升了28%和39%。 MSC光谱变换效果优于平滑+微分, 基于MSC光谱变换的PLS, LSSVM和BPNN模型检验 R²分别为0.81, 0.86和0.78, PRD分别为2.25, 2.59和2.07, 比平滑+微分提升了9%20%( R²)和11%22%(PRD), 意味着MSC具有较强的TOC信息提取能力。 GA不利于增加预测模型精度, 基于GA特征波长的模型预测 R²降低了9%36%, PRD降低了18%33%, 可能与GA提取的特征波长数量偏少有关。 BPNN模型的预测精度最低, 可能与其容易陷入局部极小点有关。 PLS模型精度较高, 可以很好的预测潮间带沉积物TOC含量。 基于异常样本剔除和MSC光谱变换, PLS模型的建模 R²为0.98, 检验 R²为0.81, RPD为2.25。 LSSVM模型精度更优于PLS, LSSVM模型建模 R²为0.99, 检验 R²和RPD分别为0.86和2.59, 显示极好的TOC定量预测能力。 总之, 针对潮间带沉积物TOC含量预测, 可以将剔除异常样本、 MSC光谱变换、 LSSVM建模结合起来, 以获得可靠、 稳定的预测模型。

Visible-near infrared spectroscopy has been shown to be a fast and effective tool for organic carbon (TOC) content prediction. However, the research target of using spectral prediction of TOC content is mainly soil or lake sediment, and there is little research on marine sediments in intertidal zone. In order to predict the content of TOC in intertidal sediments quickly and accurately, this study constructed TOC prediction model by combining abnormal sample elimination, spectral feature transformation and feature wavelength extraction, that is, collecting sediment spectra of samples in intertidal zone, using Markov distance, standard lever value and student residuals combined analysis method to remove abnormal samples, using multivariate scattering correction (MSC), smoothing + differential for spectral transformation, using genetic algorithm (GA) to extract characteristic wavelengths, using partial least squares method (PLS), least squares support vector machine (LSSVM) and BP Neural Network (BPNN) to model and predict sediment TOC content, using the decision coefficient ( R²) and residual estimation deviation (PRD) to evaluate model accuracy. The results showed that the elimination of abnormal samples improved model accuracy, and the test R² and PRD of the BPNN model increased by 28% and 39% respectively. The effect of MSC was better than that of smoothing+differential, and the test R²and PRD of PLS, LSSVM and BPNN models based on MSC were 0.81, 0.86, 0.78 and 2.25, 2.59, 2.07, respectively, which enhanced 9%20% ( R²) and 11%22% (PRD) than that based on smoothing+differential, suggesting that MSC has a strong ability to extract TOC information. GA is not conducive to increasing model accuracy, the test R² and PRD of models based on GA reducedby 9%36% and 18%33%, respectively. This may be related to the low number of characteristic wavelengths extracted by GA. The BPNN model has the lowest predictive accuracy and may be related to its vulnerability to local minimums. PLS model has high accuracy and can predict TOC content in intertidal zone. Basing on abnormal sample elimination and MSC, the modeling set R² of PLS model was 0.98, and the prediction set R² and RPD were 0.81 and 2.25 respectively. The accuracy of LSSVM model was better than that of PLS, the modeling set R² was 0.99, the test set R² and RPD were 0.86 and 2.59 respectively, implying excellent TOC quantitative prediction ability of LSSVM. In a word, for the prediction of TOC content in intertidal sediments, the combination of abnormal sample elimination, MSC spectral transformation and LSSVM modeling can obtain a reliable and stable prediction model.