| 引用本文: | 位俊燕,赵轶鸣,郝艳玲,等.湖泊悬浮物质量浓度遥感反演及变化分析
——以岱海为例[J].灌溉排水学报,2025,44(6):99-110. |
| WEI Junyan,ZHAO Yiming,HAO Yanling,et al.湖泊悬浮物质量浓度遥感反演及变化分析
——以岱海为例[J].灌溉排水学报,2025,44(6):99-110. |
|
| 摘要: |
| 【目的】对比分析传统经验模型和5种机器学习算法精度和性能，分析2017—2023年岱海悬浮物质量浓度的时空分布特征，并探讨岱海悬浮物质量浓度变化的影响因素。【方法】通过现场实验获取岱海水体实测光谱和悬浮物质量浓度，对比分析5种机器学习模型（BP神经网络模型、卷积神经网络模型、随机森林模型、支持向量机模型、径向基函数神经网络模型）和传统经验模型的遥感反演精度，选取精度最高的反演模型基于Sentinel-2 MSI卫星数据反演岱海悬浮物质量浓度。【结果】①机器学习模型较传统经验模型的反演精度高，经三折交叉和地面同步数据验证BP模型反演效果优于其他模型，模型R2为0.76，RMSE为4.66 mg/L，MAPE为26.57%；②BP模型的9个输入特征变量均对模型有正贡献作用，特征变量B4·B5对模型的重要性最为显著，其次B4、B4+B5和B5·（B4+B5）对模型精度影响较大；③2017—2023年岱海年平均悬浮物质量浓度介于8.43~11.68 mg/L，且呈小幅波动下降趋势，5—10月岱海月平均悬浮物质量浓度介于7.77~10.84 mg/L，5月悬浮物质量浓度显著高于其他月份且变化较大，6月后开始下降并趋于稳定。空间上，岱海悬浮物质量浓度主要呈“近岸高、离岸低”的特点，高值区主要分布在岱海西南和东部近岸水域；④岱海悬浮物质量浓度与风速表现出显著的正线性关系（r=0.61，p＜0.001），与气温和降水量的相关性不大。【结论】遥感技术结合机器学习模型能够大幅度提升水质监测的效率和精度。在众多模型中，BP模型和CNN模型精度优势更为明显，其中BP模型精度更高且更为稳定。 |
| 关键词: 岱海；悬浮物质量浓度；Sentinel-2 MSI；机器学习；遥感反演 |
| DOI：10.13522/j.cnki.ggps.2024342 |
| 分类号: |
| 基金项目: |
|
| Analyzing spatiotemporal variation in suspended particulate matter in lakes using remote sensing |
|
WEI Junyan, ZHAO Yiming, HAO Yanling, JIA Xiaoxue, MA Xinyan
|
|
College of Ecology and Environment, Inner Mongolia University, Hohhot 010021, China
|
| Abstract: |
| 【Objective】Suspended particulate matter (SPM) in lakes not only affects water quality but also influences ecological functions. This paper investigates the feasibility of using remote sensing to analyze the spatiotemporal variation in SPM in lakes, with an application to Daihai Lake.【Method】Field measurements were conducted to collect spectral and SPM concentration data in Daihai Lake. The accuracy of five machine learning inversion models - backpropagation (BP) neural network, convolutional neural network (CNN), random forest (RF), support vector machine (SVM), and radial basis function (RBF) neural network-was compared with that of traditional empirical models using Sentinel-2 MSI satellite data. The most accurate model was selected to estimate SPM concentrations across the lake.【Result】① Machine learning models outperformed traditional empirical models. Among them, the BP neural network achieved the highest accuracy, with a three-fold cross-validation test using synchronous ground data yielding R2=0.76, RMSE=4.66 mg/L, and MAPE=26.57%. ② All nine input variables in the BP model contributed positively to model performance. The most influential variable was the B4·B5, followed by B4, B4+B5, and B5·(B4+B5). ③ From 2017 to 2023, annual average SPM concentrations ranged from 8.43 to 11.68 mg/L, showing a slight downward trend. Monthly averages SPM concentration from May to October ranged from 7.77 to 10.84 mg/L, with May exhibiting the highest concentrations. After June, SPM levels generally declined and stabilized. Spatially, the highest concentrations occurred near the shore, particularly in the southwestern and eastern parts of the lake. ④ SPM concentrations were significantly and positively correlated with wind speed (r = 0.61, p＜ 0.001), while correlations with air temperature and precipitation were not significant.【Conclusion】Remote sensing combined with machine learning provides an effective and accurate method for monitoring SPM in lakes. This approach can help to improve management and ecological protection of lake environments. Among all models evaluated, the BP neural network is most accurate and stable, followed by the CNN model. |
| Key words: Lake Daihai; suspended particulate matter; Sentinel-2 MSI; machine learning; remote sensing inversion |
