Research On Soil Heavy Metal Content Prediction Method Based On Deep Reinforcement Learning

The prediction of spatial distribution characteristics of heavy metals in soils is an essential part of the soil pollution remediation and treatment process,but the prediction accuracy of commonly used spatial prediction algorithms in complex environments is unsatisfactory.To address this problem,this paper proposes a spatial prediction model based on deep reinforcement learning,which has higher prediction accuracy than traditional spatial prediction models and provides a basis for further improving the evaluation and treatment plan of heavy metal pollution in the study area.The main research contents are as follows:In this paper,we propose an adaptive deep Q-learning network model to address the problem that common deep reinforcement learning models have slow convergence rates in some cases and unstable states after convergence.In order to strengthen the intrinsic connection between state and action and accelerate the convergence rate of the model,the output of the value function part of the Q network is combined with the reward of the environmental feedback to form the final total reward,and the total reward is used to train the intelligence instead of the original reward.The experimental results show that the model is superior to other commonly used deep reinforcement learning models in terms of convergence rate and stability after convergence.In order to make the variogram more effectively represent the spatial structure of the current environment,a parameter estimation method for the variogram model based on deep reinforcement learning is proposed in this paper.In this paper,the parameters of the variogram model are estimated by taking advantage of the powerful representation and learning capabilities of the deep reinforcement learning model,and then a comparison experiment is conducted.The comparison experiment shows that the parameters learned from the deep reinforcement learning model are better than those computed by the common software for ground statistics(GS+).In this paper,a self-adjusting inverse distance-weighted prediction model is proposed,which firstly uses an adaptive deep Q-learning network to train and learn the hyperparameters in the inverse distance-weighted interpolation method to obtain the optimal hyperparameters for each sample point.The learned hyper-parameters are then interpolated using the Kriging method to obtain the hyperparameter distribution model.The interpolation points are then input into the hyperparameter model,and the optimal hyperparameter corresponding to the interpolation point is calculated.Finally,the information from the hyperparameter and the interpolation point is substituted into the inverse distance weighting method to obtain the predicted heavy metal content at the interpolation point.In this paper,a self-adjusted inverse distance-weighted prediction model is used to perform prediction experiments on the soil heavy metal content data set,and the experimental results show that the prediction accuracy of the proposed model improved by7.32% over the classical inverse distance weighting algorithm and by 13.03% over the random forest regression model,which was in accordance with the research expectations.