Predicting increments in heavy metal contamination in farmland soil

Traditional pollution risk assessment approaches face difficulty to effectively reflect the current state of pollution. This study aims to investigate the use of logistic regression and deep neural networks as modelling approaches to predict the probability of recontamination of irrigation areas and the pollution increment for each agricultural land in Taoyuan City, following the remediation of agricultural land pollution. The study has sampled all contaminated 2,824 parcels of farmland in Taoyuan City, involving 9,674 data entries in recent 20 years. The increment of copper, zinc, and cadmium was predicted with an accuracy of 82%, 83%, and 91%, respectively, using logistic regression and ensemble learning. In particular, the predictive power of cadmium is the best, while the deep neural network model demonstrated a good fit between the predicted copper, zinc, and cadmium increments and the actual values. The R2 values obtained were 0.81, 0.82, and 0.98 respectively, with the predictive power of cadmium being particularly significant. This indicates that the model can effectively predict the increments in heavy metals such as copper, zinc, and cadmium in the farmland and select the farmland with the highest pollution recurrence risk as the main target for soil surveys.