| 4种机器学习模型反演太湖叶绿素a浓度的比较 |
| Use of Remote Multispectral Imaging to Monitor Chlorophyll-a in Taihu Lake: A Comparison of Four Machine Learning Models |
| 投稿时间：2018-06-09 修订日期：2019-07-16 |
| DOI：10.15928/j.1674-3075.2019.04.007 |
| 中文关键词:机器学习模型 叶绿素a 太湖 |
| 英文关键词:machine learning models chlorophyll-a Taihu Lake |
| 基金项目:2017年国家重点研发计划（2017YFC0506200）；深圳市科技创新委员会基础研究学科布局项目（JCYJ20151117105543692） |
|
| 摘要点击次数: 2458 |
| 全文下载次数: 2524 |
| 中文摘要: |
| 基于太湖实测叶绿素a浓度数据以及同步HJ-1B卫星CCD多光谱影像，综合比较4种机器学习模型（随机森林，RF；支持向量回归，SVR；反向传播人工神经网络，BPANN；深度学习，DL）反演太湖叶绿素a浓度的精度、稳定性及鲁棒性。利用11种波段组合分别建立基于RF、SVR、BPANN和DL的反演模型，筛选出最佳波段组合的模型用于验证和评价。模型精度方面，DL（决定系数R2=0.91，均方根误差RMSE=3.458 μg/L，相对预测偏差RPD=3.13）和SVR（R2=0.88，RMSE=3.727 μg/L，RPD=2.90）具有较优的验证精度；模型稳定性方面，DL模型不易受模型校正样本数影响，稳定性较好，而RF模型稳定性较差；模型鲁棒性方面，DL模型不易受噪声影响，鲁棒性较好，其次是SVR、BPANN和RF模型。综合4种模型的验证精度、稳定性和鲁棒性，DL模型在太湖叶绿素a浓度的反演中具有较大应用潜力，能为其他学者利用机器学习方法研究湖泊水色参数提供借鉴。 |
| 英文摘要: |
| Lakes play a vital role in the sustainable development of human production, societies and regional economies. Most lakes in China are threatened by eutrophication, the direct cause of cyanobacteria blooms. Chlorophyll-a (Chl-a) measurements are used to indicate the degree of eutrophication and to monitor the growth and decline of harmful algal blooms (HABs). The spatial distribution of Chl-a in a water body can be used to guide the remediation and management of lake ecosystems. In this study, we compared the accuracy, stability and robustness of four machine learning models [Random Forest (RF), Support Vector Regression, (SVR), Back Propagation Artificial Neural Network (BPANN) and Deep Learning (DL)] for predicting chlorophyll-a concentration in Taihu Lake from satellite multispectral images. Calibration and validation of chlorophyll-a simulation by the four learning models were based on in-situ measurements of Chl-a concentration and synchronous HJ-1B CCD multispectral images. The spectral reflectance of 11 wave band combinations were used as model input data and the measured Chl-a concentrations were used for model calibration. The model and spectral band combination that best fit the Chl-a field data was selected as the optimum model for evaluation and validation. The calibration and validation coefficients (R2c and R2v), root mean square errors of calibration and validation (RMSEc and RMSEv), bias (Bv) and relative predictive deviation (RPD) were used to evaluate model performance. Among the four models, two displayed superior performance: DL (R2v = 0.91, RMSEv = 3.458 μg/L, RPD = 3.13) and SVR (R2v = 0.88, RMSEv = 3.727 μg/L, RPD = 2.90). The DL model was less sensitive to calibration sample size and displayed better stability than the RF model and the DL model was less sensitive to noise and more robust than the other three models. Based on the comprehensive comparison of the accuracy, stability and robustness, this study shows that the DL model has the best potential for predicting Chl-a in Taihu Lake based on multispectral imaging. Generally, the four machine learning methods, in combination with remote satellite imaging are practical for predicting Chl-a. However, incorporating water quality parameters (water temperature, dissolved oxygen, dissolved phosphorus and total nitrogen) into the model will improve the accuracy of Taihu Lake Chl-a predictions. |
| 徐 逸,董轩妍,王俊杰.2019.4种机器学习模型反演太湖叶绿素a浓度的比较[J].水生态学杂志,40(4):48-57. |
| 查看全文 查看/发表评论 下载PDF阅读器 HTML |
|
|
|
