Research on CatBoost model based on AutoEncoder dimensionality reduction in pollution source apportionment

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Water quality data is complex, high-dimensional, and non-linear.
• Traditional receptor models (PCA, PMF, APCS-MLR) struggle with these data characteristics, leading to inaccurate pollution source apportionment.
• Outliers and missing values further compromise traditional model reliability.

Purpose of the Study:

• To develop a precise pollution source apportionment model for water quality data.
• To determine the quantity, types, and contribution rates of potential pollution sources.
• To overcome the limitations of traditional receptor models in handling complex water quality data.

Main Methods:

• Principal Component Analysis (PCA) to determine the number of pollution sources.
• Autoencoder (AE) model for dimensionality reduction and pollution source identification.
• CatBoost model for quantifying the contribution rate of each identified source.
• Case study on Qinhuai New River for model validation.

Main Results:

• Identified four pollution source types: organic/domestic sewage, industrial, urban runoff/soil erosion, and agricultural.
• Quantified source contributions: organic (31.1%), industrial (21.5%), urban runoff (21.7%), and agricultural (25.7%).
• AE model achieved reconstruction R² > 0.95 and MSE < 0.05.
• CatBoost model achieved R² > 0.95 and MSE < 0.05 for contribution quantification.

Conclusions:

• The PCA-AE-CatBoost model provides a more accurate technical basis for water pollution control.
• This advanced model outperforms PCA-APCS-MLR and PCA-CatBoost models in accuracy and reliability.
• The findings offer crucial insights for targeted water quality management strategies.