Inversion of Pb content in soil based on iron oxide characteristic spectrum and improved genetic algorithm

Abstract: A hyperspectral reflectance can be expected as an alternative method to quickly predict the concentration of heavy metals in soil at present. However, excessive band counting of hyperspectral data can lead to information redundancy during the quantitative inversion of surface parameters, and thereby to make it difficult to extract valid data from the spectrum. In this study, two aspects were addressed to solve spectral redundancy during the inversion of Pb content in soil: One is to extract the characteristics of iron oxides from the soil spectrum, according to the adsorption mechanism of Pb element in soil, in order to reduce data redundancy, while enhance the mechanism of inversion. Another is to propose an Improved Genetic Algorithm (IGA), in order to deal with the premature convergence of traditional Genetic Algorithm (GA), further to improve the extracting ability of algorithm for the required spectrum. Taking the Xiong'an New Area near Beijing, a national new-district in China, as the research region, the field experiment was carried out to quickly collect the Pb content of soil in the farmland, resulting from the rapid construction and developing facilities in recent years. 70 soil samples from the Xiong'an farmland were selected to establish the inversion model of Pb content. SVC HR1024I type ground object spectrometer was used to measure the raw hyperspectral reflectance of soil samples according to the national standard procedure in field farmland. After that, the collected soil samples were transferred to the laboratory for chemical analysis, particularly for the Pb content. Before modeling, the methods, including Savitzky-Golay filter, envelope removal and first-order difference, were used to pretreat with segmented the curves of soil reflectance. Partial Least Squares Regression (PLSR) was selected to serve as inversion model, in order to well deal with the strong correlation between hyperspectral bands. The models were also constructed in terms of full spectrum and characteristic bands of iron oxides, to verify the effectiveness of extracting method from characteristic bands of iron oxides during the inversion process. In full spectrum, the average values of R2, RPD, RMSE in the IGA-PLSR models were 0.440, 1.366, and 4.012 mg/kg, respectively. In the characteristic bands of iron oxides, the average values of R2, RPD, RMSE in the IGA-PLSR models were 0.822, 2.377, and 2.221 mg/kg, respectively. It infers that the models using the characteristic bands of iron oxides were much better than those using the full spectrum, indicating that the extraction for the characteristic bands of iron oxides can be contributed to reduce the redundancy of spectral information, and thereby to significantly improve the accuracy of the inversion model. Consequently, a final inversion model was selected from the optimal model for the improvement of inversion algorithm, based on the utilization of the characteristic bands of iron oxides to construct the GA-PLSR and IGA-PLSR models with different iterations. In the GA-PLSR models, the average values of R2, RPD, RMSE in the optimal model were 0.788, 2.140, and 2.447 mg/kg, respectively, whereas, the IGA-PLSR model achieved the highest accuracy, where the average values of R2, RPD, RMSE were 0.837, 2.403, and 2.063 mg/kg, respectively. Therefore, the optimal model of IGA-PLSR was chosen to estimate the Pb content in soil for the Xiong'an farmland area. The IGA-PLSR method has demonstrated better performance in the inversion of Pb content in soil, indicating that the proposed method was valid to improve genetic algorithm. At last, a visualized profile can be conducted for the changes of RMSE during the 1000 iterations of two models. The visualization revealed that the RMSE in the GA-PLSR model decreased monotonously during the operation process, whereas, the RMSE in the IGA-PLSR model fluctuated greatly at the beginning of the iteration, showing a certain probability to jump out from the local solution area to a better solution area at the later stage, while meaning that the problem of "premature convergence" was solved properly. This study revealed that two aspects, including the extraction of iron oxides spectral features associated with Pb absorption, and the improved algorithm of bands selection, both can be used to enhance the accuracy of Pb content estimation in soil. The proposed algorithm can provide a sound reference to real-time monitor the variation of heavy metal content in soil for Xiong'an area, China.