With the development of industrialization, environmental heavy metal pollution has become increasingly serious, and the growth of crops has been seriously affected by heavy metal pollution in the soil environment. Therefore, it is necessary to establish methods for distinguishing and monitoring heavy metal pollution. The application of hyperspectral remote sensing in heavy metal pollution monitoring demonstrates the great potential of using crop leaf spectra to accurately distinguish heavy metal pollution elements. At the same time, new spectral processing methods and models are required to provide support for accurate identification. In this study, greenhouse experiments were conducted to simulate the growth of corn plants under heavy metal Cu and Pb stress. Collect hyperspectral data from different leaf layers of maize plants during the heading stage. Multivariate empirical mode decomposition (MEMD) was introduced, and the spectral data were preprocessed using MEMD, First derivative (FD), and second derivative (SD). At the same time, chemical analysis was used to examine the changes in copper (Cu), lead (Pb), and chlorophyll content in corn leaves. Competitive adaptive reweighted sampling (CARS) and iteratively retaining informative variables (IRIV) were used to screen characteristic bands that were sensitive to copper and lead. Finally, machine learning SVM, ELM, and XGBoost were utilized to construct and propose a series of models such as MEMD-CARS-ELM for accurate discrimination of Cu and Pb pollution elements. The results indicated that the discriminative model established after the MEMD transformation of the spectrum exhibited the best performance. Among them, whether it is tender leaves, functional leaves, or basal leaves, the accuracy of the MEMD-CARS-SVM and MEMD-CARS-ELM models in the training group and validation group for distinguishing Cu and Pb elements is greater than 80%. Other models established by MEMD spectral transformation are also significantly better at identifying Cu and Pb than those established by FD and SD transformations. The signal time–frequency analysis method MEMD is feasible and excellent for hyperspectral data processing. Based on this method, the Cu and Pb pollution element identification method proposed in this study was reliable. The research results provide a new method for the preprocessing of hyperspectral data and a new perspective for the accurate identification of soil heavy metal contamination elements. This study showed that corn leaf spectra can be used to accurately identify heavy metal pollution elements, providing a powerful scientific reference for hyperspectral remote sensing to monitor heavy metal pollution in large areas.
Read full abstract