-
浊度 (turbidity,T,NTU) 是水环境监测的重要参数,反映了光在水体传输中的受阻程度,浊度的变化能够表征水中悬浮颗粒物浓度的变化,对水中污染物质迁移具有重要指示作用[1-2],浊度还会通过影响水下光的传播改变水生生态系统的初级生产力和生态平衡[3],因此浊度对于水资源管理、水生态保护等具有重要意义。长三角地区水环境状况复杂多变,前人对长三角地区的水体浊度[4]、总悬浮物浓度 (total suspended matter concentration,TSM,mg∙L−1) [5]、透明度[6]等开展了众多研究,但研究多为大型河湖水体,中小型湖泊和河流研究较少,长三角示范区是引领长三角地区高质量一体化发展的重要区域,浊度反演研究可为示范区绿色发展提供重要依据。
如何利用卫星数据准确估算水体浊度是国内外学者研究的重点,常见的有经验模型、半经验模型和机器学习模型。HOU等[5]运用MODIS 555 nm和645 nm波段表面反射率比值构建了总悬浮物浓度经验模型用于长江中下游流域大型湖泊和水库的浊度动态监测。DOXARAN等[7]基于MODIS 858 nm和645 nm波段表面反射率比值构建浊度反演经验模型,用于研究法国吉伦特河口最大浑浊带浊度动态变化。DOGLIOTTI等[8]从固有光学特性出发基于MODIS 645 nm波段遥感反射率 (remote sensing reflectance,Rrs,sr−1) 发展了适用于沿海和河口水域中低浊度的半经验反演模型。MA等[9]运用梯度增强决策树 (gradient boosting decision trees,GBDT) 估算了中国东北地区湖泊浊度。前人研究的浊度反演模型大多应用于开阔的近海河口水域,受人类活动影响较小,研究对象多为光学敏感性较高的高度浑浊水体。而本研究水体主要为中低浊度水体,且研究区内湖荡密布,航道众多,生产生活密集,水动力变化复杂,针对此类水体浊度的长时序监测研究尚不多见。
本研究以长三角示范区内典型河湖为研究区,运用实测数据构建并对比了经验模型、半经验模型和机器学习模型三类浊度反演模型,随后基于最佳的XGBoost (eXtreme Gradient Boosting) 模型运用Landsat卫星数据反演研究区浊度,分析浊度的长期时空动态变化及其驱动因素。
基于XGBoost的内陆河湖浊度反演与长时序分析
Inversion and long-term series analysis of turbidity in inland rivers and lakes based on XGBoost
-
摘要: 浊度是影响水下光场及营养盐循环的关键要素之一,浊度监测可以为河湖水质的污染防控和预警提供科学依据。以长三角示范区的典型河湖为研究区,使用实测数据构建浊度反演模型,并利用1984—2022年Landsat卫星数据分析了研究区河湖浊度的长时序动态变化。对比传统经验模型、半经验模型和机器学习模型,XGBoost机器学习模型精度最高 (R2为0.68,RMSE为4.78 NTU) 。浊度反演结果表明,近40年河流航道和淀山湖北部非渔场区域浊度上升了10%和12%,而元荡湖和大莲湖浊度下降了19%和27%,并且浊度随着建设用地面积的增加而增大;研究区浊度季节性变化显著,秋冬季平均浊度比春夏季高6 NTU,月平均浊度与月平均降水量负相关,相关系数为-0.61 (p<0.05) ,但与月平均风速没有显著的相关性。基于XGBoost的Landsat长时序浊度反演能够把握研究区浊度的时空变化趋势,明确水污染管理与治理方向,最终助力长三角一体化发展。Abstract: Turbidity is one of the key elements affecting the underwater light field and nutrient cycling. Turbidity monitoring can provide a scientific basis for pollution prevention, control, and early warning of river and lake water quality. The typical rivers and lakes in the demonstration zone of the Yangtze River Delta were taken as the study area. The turbidity inversion model was constructed using in-situ data, based on which the long-term dynamic changes of turbidity in the rivers and lakes of the study area were analyzed using a total of 323 Landsat TM/ETM+/OLI images from 1984 to 2022. Through the comparison between the traditional empirical model, semi-empirical model, and machine learning model, the machine learning model named XGBoost demonstrated the highest accuracy (R2 and RMSE were 0.68 and 4.78 NTU, respectively). The results of turbidity inversion showed that, in the last 40 years, turbidity in the river channel and the northern non-fishing area of Dianshan Lake increased by 10% and 12%, respectively, while turbidity in Yuandang Lake and Dalian Lake decreased by 19% and 27%, respectively. Moreover, it was found that turbidity increased with the expansion of the built-up land area. The seasonal variation of turbidity in the study area was significant and the average turbidity in autumn and winter was 6 NTU higher than that in spring and summer. The monthly average turbidity was negatively correlated with the monthly average precipitation (r=-0.61, p<0.05), but its correlation with the monthly average wind speed was found to be insignificant. The XGBoost-based long-term inversion of turbidity from Landsat images can not only capture the spatiotemporal trend of turbidity in the study area, but also reveal the direction of water pollution management and control, eventually contributing to the integrated development of the Yangtze River Delta.
-
Key words:
- turbidity /
- XGBoost /
- Landsat /
- long-term series /
- demonstration zone of the Yangtze River Delta
-
表 1 采样位置、时间和数量
Table 1. Sampling location, time and number of samples
编号 位置 时间 采样数量 有效数据量 1 太湖西山岛以北 2022-10-24 35 6 2 元荡湖 2022-10-25 35 29 3 太浦河、汾湖、三白荡 2022-11-23 18 15 4 淀山湖南部 2022-12-13 20 15 5 淀山湖、元荡湖 2023-03-26 18 16 表 2 基于QAA计算OLI的颗粒物后向散射系数bbp(655)
Table 2. Calculation of bbp at 655 nm for OLI based on QAA
步骤 计算 1 $ {r_{rs}}(\lambda ) = {R_{rs}}(\lambda )/(0.52+1.7{R_{rs}}(\lambda )) $ 2 , 其中$ \mu (\lambda ) = \dfrac{{ - {g_0}+\sqrt {{{({g_0})}^2}+4{g_1} \times {r_{rs}}(\lambda )} }}{{2{g_1}}} $ = 0.089,$ {g_0} $ = 0.1245$ {g_1} $ 如果Rrs(655)<0.0015 sr−1 否则655 nm = $ {\lambda _0} $ 3 $ \chi = log\left( {\dfrac{{{r_{rs}}(443)+{r_{rs}}(483)}}{{{r_{rs}}(561)+5{r_{rs}}(655)/{r_{rs}}(483) \times {r_{rs}}(655)}}} \right) $ $ a({\lambda _0}) = a(561) = {a_{\text{w}}}(561)+{10^{{h_0}+{h_1}\chi+{h_2}{\chi ^2}}} $ ,$ {h_0} = - 1.146 $ ,$ {h_1} = - 1.366 $ $ {h_2} = - 0.469 $ $ a({\lambda _0}) = a(655)\; = {a_{\text{w}}}(655)+0.39{(\dfrac{{{R_{rs}}(655)}}{{{R_{rs}}(443)+{R_{rs}}(483)}})^{1.14}} $ 4 $ {b_{{\text{bp}}}}({\lambda _0}) = {b_{{\text{bp}}}}(561) = \dfrac{{\mu (561) \times a(561)}}{{1 - \mu (561)}} - {b_{{\text{bw}}}}(561) $ $ {b_{{\text{bp}}}}({\lambda _0}) = {b_{{\text{bp}}}}(655) = \dfrac{{\mu (655) \times a(655)}}{{1 - \mu (655)}} - {b_{{\text{bw}}}}(655) $ 5 $ \eta = 2.0(1 - 1.2exp( - 0.9{r_{rs}}(443)/{r_{rs}}(561))) $ 6 $ {b_{{\text{bp}}}}(\lambda ) = {b_{{\text{bp}}}}({\lambda _0}){({\lambda _0}/\lambda )^\eta } $ 注:表中a和aw分别为总吸收系数和纯海水吸收系数,m−1;bbw和bbp分别为纯海水和悬浮颗粒的后向散射系数,m−1。rrs(λ)为水表面以下遥感反射率,sr−1,Rrs(λ)是水面以上遥感反射率,sr−1。 表 3 遥感反射率光谱特征组合与浊度的相关性
Table 3. Correlation between combinations of Rrs spectral features and turbidity
波段组合 相关系数r 波段组合 相关系数r B2 0.49** B3 0.50** B4 0.71** B5 0.65** B4/B2 0.78** B4/B3 0.88** B4/B5 −0.23* (B4-B2)/(B4+B2) 0.76** (B4-B3)/(B4+B3) 0.87** (B4-B5)/(B4+B5) −0.38** 注:*表示相关性在0.05水平上显著;**表示相关性在0.01水平上显著。 表 4 浊度反演模型参数化及比较
Table 4. Parameterization and comparison of turbidity inversion models
序号 模型 表达式 留一交叉验证 R2 RMSE (NTU) MAPE (%) 1 二次多项式 $ y = 136.38{x_1}^2 - 97.36{x_1}+19.55 $ 0.78 5.25 31.43 2 指数函数 $ y = 5{e^{2.72{x_1}}} - 15.16 $ 0.78 5.25 30.35 3 幂函数 $ y = 62.7{x_1}^{2.99} - 3.28 $ 0.78 5.26 30.93 4 SESB-based $ y = 117.14{\rho _{\text{w}}}(\lambda )/(1 - {\rho _{\text{w}}}(\lambda )/0.1686)+9.63 $ 0.39 8.72 57.91 5 QAA-based $ y = 74.14x_2^{0.96} $ 0.54 7.60 48.83 6 XGBoost $ y = {f_{{\text{XGBoost}}}}({x_1}) $ 0.83 4.58 23.79 注:x1表示OLI传感器B4/B3,x2表示bbp(655),m−1,y表示实测浊度,NTU;fXGBoost表示XGBoost模型函数。 -
[1] PETUS C, CHUST G, GOHIN F, et al. Estimating turbidity and total suspended matter in the Adour River plume (South Bay of Biscay) using MODIS 250-m imagery[J]. Continental Shelf Research, 2010, 30(5): 379-392. doi: 10.1016/j.csr.2009.12.007 [2] NECHAD B, RUDDICK K G, PARK Y. Calibration and validation of a generic multisensor algorithm for mapping of total suspended matter in turbid waters[J]. Remote Sensing of Environment, 2010, 114(4): 854-866. doi: 10.1016/j.rse.2009.11.022 [3] DING W H, ZHAO J X, QIN B Q, et al. Exploring and quantifying the relationship between instantaneous wind speed and turbidity in a large shallow lake: case study of Lake Taihu in China[J]. Environmental Science and Pollution Research, 2021, 28(13): 16616-16632. doi: 10.1007/s11356-020-11544-y [4] LIN X N, WU M, SHAO X X, et al. Water turbidity dynamics using random forest in the Yangtze River Delta Region, China[J]. Science of the Total Environment, 2023, 903: 166511. doi: 10.1016/j.scitotenv.2023.166511 [5] HOU X J, FENG L, DUAN H T, et al. Fifteen-year monitoring of the turbidity dynamics in large lakes and reservoirs in the middle and lower basin of the Yangtze River, China[J]. Remote Sensing of Environment, 2017, 190: 107-121. doi: 10.1016/j.rse.2016.12.006 [6] YIN Z Y, LI J S, LIU Y, et al. Water clarity changes in Lake Taihu over 36 years based on Landsat TM and OLI observations[J]. International Journal of Applied Earth Observation and Geoinformation, 2021, 102: 102457. doi: 10.1016/j.jag.2021.102457 [7] DOXARAN D, FROIDEFOND J, CASTAING P, et al. Dynamics of the turbidity maximum zone in a macrotidal estuary (the Gironde, France): Observations from field and MODIS satellite data[J]. Estuarine, Coastal and Shelf Science, 2009, 81(3): 321-332. doi: 10.1016/j.ecss.2008.11.013 [8] DOGLIOTTI A I, RUDDICK K G, NECHAD B, et al. A single algorithm to retrieve turbidity from remotely-sensed data in all coastal and estuarine waters[J]. Remote Sensing of Environment, 2015, 156: 157-168. doi: 10.1016/j.rse.2014.09.020 [9] MA Y, SONG K S, WEN Z D, et al. Remote sensing of turbidity for lakes in northeast China using Sentinel-2 images with machine learning algorithms[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14: 9132-9146. doi: 10.1109/JSTARS.2021.3109292 [10] VANHELLEMONT Q, RUDDICK K. Acolite for Sentinel-2: Aquatic applications of MSI imagery[C]// Proceedings of the 2016 ESA Living Planet Symposium. Prague: 2016: 9-13. [11] VANHELLEMONT Q. Sensitivity analysis of the dark spectrum fitting atmospheric correction for metre- and decametre-scale satellite imagery using autonomous hyperspectral radiometry[J]. Optics Express, 2020, 28(20): 29948. doi: 10.1364/OE.397456 [12] VANHELLEMONT Q. Adaptation of the dark spectrum fitting atmospheric correction for aquatic applications of the Landsat and Sentinel-2 archives[J]. Remote Sensing of Environment, 2019, 225: 175-192. doi: 10.1016/j.rse.2019.03.010 [13] VANHELLEMONT Q, RUDDICK K. Atmospheric correction of metre-scale optical satellite data for inland and coastal water applications[J]. Remote Sensing of Environment, 2018, 216: 586-597. doi: 10.1016/j.rse.2018.07.015 [14] TANG R G, SHEN F, PAN Y Q, et al. Multi-source high-resolution satellite products in Yangtze Estuary: cross-comparisons and impacts of signal-to-noise ratio and spatial resolution[J]. Optics Express, 2019, 27(5): 6426. doi: 10.1364/OE.27.006426 [15] LI P, KE Y H, BAI J H, et al. Spatiotemporal dynamics of suspended particulate matter in the Yellow River Estuary, China during the past two decades based on time-series Landsat and Sentinel-2 data[J]. Marine Pollution Bulletin, 2019, 149: 110518. doi: 10.1016/j.marpolbul.2019.110518 [16] LUO W, SHEN F, HE Q, et al. Changes in suspended sediments in the Yangtze River Estuary from 1984 to 2020: Responses to basin and estuarine engineering constructions[J]. Science of the Total Environment, 2022, 805: 150381. doi: 10.1016/j.scitotenv.2021.150381 [17] VANHELLEMONT Q. Daily metre-scale mapping of water turbidity using CubeSat imagery[J]. Optics Express, 2019, 27(20): A1372-A1399. doi: 10.1364/OE.27.0A1372 [18] 唐军武, 田国良, 汪小勇, 等. 水体光谱测量与分析Ⅰ: 水面以上测量法[J]. 遥感学报, 2004(1): 37-44. doi: 10.11834/jrs.20040106 [19] 国家环境保护总局. 地表水和污水监测技术规范: HJ/T 91-2002[S]. 北京: 中国环境出版社, 2002. [20] 国家环境保护局. 水质 悬浮物的测定 重量法: GB 11901-89[S]. 1990. [21] LUO W, LI R H, SHEN F, et al. HY-1C/D CZI image atmospheric correction and quantifying suspended particulate matter[J]. Remote Sensing, 2023, 15(2): 386. doi: 10.3390/rs15020386 [22] SHI K, ZHANG Y L, ZHU G W, et al. Long-term remote monitoring of total suspended matter concentration in Lake Taihu using 250 m MODIS-Aqua data[J]. Remote Sensing of Environment, 2015, 164: 43. doi: 10.1016/j.rse.2015.02.029 [23] ALLAM M, YAWAR ALI KHAN M, MENG Q. Retrieval of turbidity on a spatio-temporal scale using Landsat 8 SR: A case study of the Ramganga River in the Ganges Basin, India[J]. Applied Sciences, 2020, 10(11): 3702. doi: 10.3390/app10113702 [24] KRATZER S, KYRYLIUK D, EDMAN M, et al. Synergy of satellite, in situ and modelled data for addressing the scarcity of water quality information for eutrophication assessment and monitoring of Swedish Coastal Waters[J]. Remote Sensing, 2019, 11(17): 2051. doi: 10.3390/rs11172051 [25] LEE Z P, CARDER K L, ARNONE R A. Deriving inherent optical properties from water color: a multiband quasi-analytical algorithm for optically deep waters[J]. Applied Optics, 2002, 41(27): 5755-5772. doi: 10.1364/AO.41.005755 [26] BALASUBRAMANIAN S V, PAHLEVAN N, SMITH B, et al. Robust algorithm for estimating total suspended solids (TSS) in inland and nearshore coastal waters[J]. Remote Sensing of Environment, 2020, 246: 111768. doi: 10.1016/j.rse.2020.111768 [27] LIU H Z, LI Q Q, BAI Y, et al. Improving satellite retrieval of oceanic particulate organic carbon concentrations using machine learning methods[J]. Remote Sensing of Environment, 2021, 256: 112316. doi: 10.1016/j.rse.2021.112316 [28] GHATKAR J G, SINGH R K, SHANMUGAM P. Classification of algal bloom species from remote sensing data using an extreme gradient boosted decision tree model[J]. International Journal of Remote Sensing, 2019, 40(24): 9412-9438. doi: 10.1080/01431161.2019.1633696 [29] CAO Z G, MA R H, DUAN H T, et al. A machine learning approach to estimate chlorophyll-a from Landsat-8 measurements in inland lakes[J]. Remote Sensing of Environment, 2020, 248: 111974. doi: 10.1016/j.rse.2020.111974 [30] CHEN T Q, GUESTRIN C. XGBoost: a scalable tree boosting system[C]// Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: Association for Computing Machinery, 2016: 785-794. [31] NECHAD B, RUDDICK K G, NEUKERMANS G. Calibration and validation of a generic multisensor algorithm for mapping of turbidity in coastal waters[C]// Remote Sensing of the Ocean, Sea Ice, and Large Water Regions 2009. SPIE, 2009: 7473: 74730H. [32] CASTAGNA A, AMADEI MARTÍNEZ L, BOGORAD M, et al. Optical and biogeochemical properties of diverse Belgian inland and coastal waters[J]. Earth System Science Data, 2022, 14(6): 2697-2719. doi: 10.5194/essd-14-2697-2022 [33] WONG T T. Performance evaluation of classification algorithms by k-fold and leave-one-out cross validation[J]. Pattern Recognition, 2015, 48(9): 2839-2846. doi: 10.1016/j.patcog.2015.03.009 [34] ZHENG Z B, LI Y M, GUO Y L, et al. Landsat-based long-term monitoring of total suspended matter concentration pattern change in the wet season for Dongting Lake, China[J]. Remote Sensing, 2015, 7(10): 13975-13999. doi: 10.3390/rs71013975