基于PCA-MLR和PCA-BPN的莱州湾南岸滨海平原土壤有机质高光谱预测研究

土壤有机质(SOM)含量是衡量土壤质量高低的重要指标，可以用高光谱快速测定。在以往研究中，估算模型多以特征波段与线性经验模型为基础进行构建，较少考虑波段间信息冗余和共线性，预测效果不很理想并难以进行推广。为最大化消除波段信息噪声，提高模型预测精度，选取莱州湾南岸滨海平原为研究区，系统采集了111个土壤样本和实测高光谱数据(325～1 075 nm)，并测试了土壤样本的有机质含量作为因变量；通过主成分分析(PCA)将实测光谱信息降维为6个主成分，并提取水分、植被光谱特征指数(DI)，以此作为自变量；最后建立多元逐步线性回归(MLR)和BP神经网络(BPN)预测模型，分析不同模型对土壤有机质预测的效果。结果表明：①经过主成分的波段信息分析判别提取出6个主成分，可以表征叶绿素残留物、盐分、腐殖酸、物化矿渣和微地貌的光谱特征。②基于6个主成分作为自变量所建立的BPN模型预测精度优于MLR模型，他们的R2分别为0.704和0.643。将水分和植被光谱特征指数作为自变量增加到预测模型后，MLR和BPN的预测精度分别提高了6.1%和5.2%，R2达到0.712和0.764；③将光谱主成分和光谱特征指数作为自变量的BPN模型进行土壤有机质预测可得到精度较高的预测结果，在土壤有机质的预测与制图中具有一定的应用潜力。

Prediction of Soil Organic Matter Based PCA-MLR and PCA-BPN Algorithm Using Field VNIR Spectroscopy in Coastal Soils of Southern Laizhou Bay

Soil organic matter (SOM) content is an important indicator of soil quality, which could be predicted using hyper spectral data rapidly. A total of 111 soil samples and hyperspectral data (325~1 075 nm) were collected from the coastal plain on the southern of Laizhou Bay. In previous study, most prediction model is based on the characteristic band and the linear empirical model, ignoring the information redundancy and collinearity between bands, the prediction accuracy is not high and it is difficult to be extended to other regions. In order to maximize the elimination of band information noise and improve the model prediction accuracy,the organic matter content of soil samples was measured as dependent variable. Through principal component analysis (PCA), the measured spectral is reduced to 6 principal components, and water and vegetation spectral characteristic indices are extracted as the independent variables. At last, we analyzed the prediction effect of different model on soil organic matter with MLR and BPN models. The results show: (1) the six principal components extracted from PCA on spectral information could be used to characterize the spectral characteristics of chlorophyll, salt, humic acid, materialized slag and micro- Landform. (2) The prediction accuracy of BPN model based on 6 principal components as independent variable is better than that of MLR model with R2 of 0.704 and 0.643 respectively. After adding water and vegetation spectral characteristic index as an independent variable to prediction models, the prediction accuracy of MLR and BPN increased by 6.1% and 5.2%, and R2 reached 0.712 and 0.764, respectively; (3) BPN model based spectral principal components and spectral characteristic indices as independent variables could predict soil organic matter with the highest accuracy, which has a potential application in soil organic matter prediction and mapping.