Water resources serve as the cornerstone of human livelihoods and economic
progress, with intrinsic links to both public health and environmental
well-being. The accurate prediction of water quality stands as a pivotal factor
in enhancing water resource management and combating pollution. This research,
employing diverse performance metrics, assesses the efficacy of five distinct
models, namely, linear regression, Random Forest, XGBoost, LightGBM, and MLP
neural network, in forecasting pH values within Georgia, USA. Concurrently,
LightGBM attains the highest average precision among all models examined.
Tree-based models underscore their supremacy in addressing regression
challenges. Furthermore, the performance of MLP neural network is sensitive to
feature scaling. Additionally, we expound upon and dissect the reasons behind
the superior precision of the machine learning models when they are compared to
the original study, which factors in time dependencies and spatial
considerations. The primary objective of this endeavor is to establish a robust
predictive pipeline, specifically tailored for practical applications. It
caters not only to individuals well-versed in the realm of data science but
also to those lacking specialization in particular application domains. In
essence, we offer a fresh perspective for achieving relative precision in data
science methodologies, emphasizing both prediction accuracy and
interpretability.

利用不同性能度量评估线性回归、随机森林、XGBoost、LightGBM 和 MLP 神经网络五种模型在美国佐治亚州预测 pH 值方面的效果，表明 LightGBM 具有最高的平均精度，树模型在回归问题中具有卓越性能，MLP 神经网络的性能对特征缩放敏感，并解释了机器学习模型相对原研究的提升原因，以建立适用于实际应用的强大预测流程，既适合数据科学领域专业人士，也适用于缺乏特定应用领域专业知识的人群，强调预测准确性和可解释性。