In Environmental science and pollution research international
Estimating heavy metal concentrations in soil-rice systems is of great significance to identify the factors controlling heavy metal transfer in soil-crop ecosystems. Recent research utilizes the advantage of convolutional calculations to extract and learn complicated information from 17 environmental covariates in rice and achieve promising results. However, as the complexity and interconnectivity in soil-crop ecosystem, just relying on convolutional calculations and a deep network structure is far from enough. The data processed by traditional deep learning technologies even with convolutional calculations are limited to Euclidean space; these architectures do not have the ability to extract information from the relationships in graph structures, which may contain rich information. Thus, in this paper, we try to integrate graph information into convolutional calculations for heavy metal prediction and propose a model named ConvGNN-HM. ConvGNN-HM combines the advantages of graph learning and convolutional calculations to predict heavy metal concentrations in a soil-rice system with analysis of 17 environmental factors. For comparison, we conduct an experiment to compare ConvGNN-HM with techniques with convolutional neural networks, multilayer perceptron, back-propagation neural networks, support vector machine, random forest, Bayesian ridge regression, and multiple linear regression. The experimental results illustrate that ConvGNN-HM got the best prediction values; the R2 values of ConvGNN-HM for cadmium (Cd), plumbum (Pb), chromium (Cr), arsenic (As), and hydrargyrum (Hg) in rice were 0.84, 0.75, 0.79, 0.49, and 0.83, respectively, and the MAE values were also acceptable. We further conduct sensitivity analysis to demonstrate the stability and robustness of ConvGNN-HM. This study demonstrates the usefulness of combining graph learning and convolutional calculations in the prediction of heavy metal concentrations and provides a new perspective to build multidimensional and multi-scale complex ecosystem models.
Zhang Zhuo, Li Yuanyuan, Bai Yang, Li Ya, Liu Meng
2023-Jan-23
Comparison, Environmental factor, Machine learning, Modeling, Sensitivity analysis
