基于Levy飞行和麻雀搜索算法优化集成学习模型的水质估算

Estimation of Water Quality Parameters Using an Ensemble Learning Model Optimized with Levy Flight and Sparrow Search Algorithms

摘要:

由于水体的光学复杂性和不同水质参数之间的相互作用，利用集成学习方法估算水质参数具有优势；然而，在建模过程中如何合理选择超参数仍然是一个难题。麻雀搜索算法能够快速搜索集成学习模型的最优参数；而Levy飞行算法可以防止麻雀搜索算法（Sparrow Search Algorithm，SSA）陷入局部最优，并提高模型的准确性和效率。使用Levy飞行算法和麻雀搜索算法对随机森林（Random Forest，RF）、自适应回归（AdaBoost Regression，ABR）和类别提升回归（CatBoost Regression，CBR）3种集成学习模型进行了优化。以郑州东风渠和熊耳河为研究区 ，基于实测叶绿素a（chlorophyll-a，Chl-a）和总悬浮物（total suspended solids，TSM）数据，构建了LSSA-RF、LSSA-ABR和LSSA-CBR这3种估算模型。实验结果表明：模型经过优化后，各项指标均有不同程度的提高。其中表现最优的是LSSA-CBR模型；CBR模型是在梯度提升框架下进行的建模，对比RF和CBR模型具有更高维度的学习能力。在叶绿素a的估算中，LSSA-CBR估算模型的均方根误差为2.325 μg·L^-1，决定系数为0.896；在总悬浮物的估算中，LSSA-CBR模型的均方根误差为1.598 mg·L^-1，决定系数为0.882。最后，将精度较好的LSSA-CBR模型应用于卫星Planet影像中，以评估河流叶绿素a和总悬浮物的空间分布情况。研究结果可为环保部门快速了解城市河流水质分布及进行水质评价与管理提供参考。

Abstract:

Due to the optical complexity of water bodies and the interactions among various water quality parameters， utilizing ensemble machine learning methods for estimating water quality parameters offers advantages. However， selecting hyperparameters in the modeling process remains challenging. The sparrow search algorithm （SSA） can rapidly search for optimal parameters of ensemble machine learning models， while the Levy flight algorithm prevents SSA from being trapped in local optima， thereby improving the accuracy and efficiency of the model. In this paper， the Levy flight algorithm and SSA were used to optimize three ensemble learning models： random forest （RF）， AdaBoost regression （ABR）， and CatBoost regression （CBR）. Taking Zhengzhou Dongfeng Canal and Xiong’er River as the study area， estimation models （LSSA-RF， LSSA-ABR， and LSSA-CBR） were developed based on measured chlorophyll-a and total suspended solids concentrations. The experimental results show that after optimization， various indicators show improvements to varying degrees. Among them， the LSSA-CBR model exhibits the best performance. The CBR model， which is modeled under the gradient boosting framework， demonstrates higher learning capability compared to RF and ABR models. For the estimation of chlorophyll-a， the root mean square error （RMSE） of the LSSA-CBR estimation model is 2.325 μg·L^-1， and the coefficient of determination （R²） is 0.896. For the estimation of total suspended solids， the RMSE of the LSSA-CBR model is 1.598 mg·L^-1， and R² is 0.882. Finally， the LSSA-CBR model， demonstrating strong accuracy， was applied to Planet images to evaluate the spatial distribution of chlorophyll-a and total suspended solids in rivers， providing a valuable reference for quickly understanding the distribution of urban river water quality and conducting water quality assessment and management.