Remote estimation of chlorophyll-a in turbid inland waters: Three-band model versus GA-PLS model
Remote estimation of chlorophyll-a in turbid inland waters: Three-band model versus GA-PLS model
复制标题
DOI：
10.1016/j.rse.2013.05.017
复制
发表时间：
2013-09
影响因子：
13.5
通讯作者：
K. Song;Lin Li;L. Tedesco;Shuai Li;H. Duan;Dianwei Liu;Bob Hall;Jia Du;Zuchuan Li;Kun Shi
中科院分区：
文献类型：
--
作者：
K. Song;Lin Li;L. Tedesco;Shuai Li;H. Duan;Dianwei Liu;Bob Hall;Jia Du;Zuchuan Li;Kun Shi
内陆和沿海浑浊沃茨叶绿素a浓度的精确遥感反演是一项具有挑战性的任务，由于其光学复杂性。基于遗传算法选择光谱变量和偏最小二乘（GA-PLS）回归的优点，建立了自适应模型。本文件的目标是：（1）利用美国中部印第安纳州（CIN）、南澳（SA）、中国东部太湖（THL）和中国东北石头口门水库（STKR）9个水体的1140个采样站的Chl-a和悬浮泥沙数据，与公认的三波段模型进行比较，评价GA-PLS模型的性能。（2）用模拟的ESA/Sentinel 3/OLCI和ESPRIT光谱评价GA-PLS的空间可移植性。GA-PLS和三波段模型对SA数据集的校正（Cal）结果准确，R2在0.98以上，相应的验证（瓦尔）结果表明，窄带光谱的相对均方根误差（rRMSE）小于6.2%。GA-PLS和三带模型均显示CIN数据集（Cal：R2=0.91和0.77;瓦尔：rRMSE=20.1%和33.4%）、THL数据集（Cal：R2=0.91和0.88;瓦尔：rRMSE=30.1%和33.7%）和STKR数据集（R2=0.84和0.82; rRMSE=29.1%和33.2%）的稳定性能。结果还表明，模拟OLCI数据集降低了GA-PLS的性能，特别是三波段模型的性能，由于粗糙和不连续的光谱配置。相比之下，GA-PLS和三波段模型都显示出改进的结果与模拟的数据集。我们的观察表明，GA-PLS模型优于三波段模型的空间可转移性，但三波段模型有自己的优点，考虑到它的简单性。进一步的分析表明，光谱测量协议，仪器和无机悬浮物影响GA-PLS和三波段模型的性能。
Accurate remote retrieval of chlorophyll-a (Chl-a) concentrations for inland and coastal turbid waters is a challenging task due to their optical complexity. An adaptive model was developed based on the merits of coupling a genetic algorithm to select spectral variables and partial least squares (GA-PLS) for regression. The objectives of this paper are: (1) to evaluate the GA-PLS model performance using datasets collected from 1140 stations encompassing a wide range of Chl-a and suspended sediment from nine water bodies across Central Indiana (CIN), USA, South Australia (SA), Taihu Lake (THL) in East China and Shitoukoumen Reservoir (STKR) in Northeast China with comparison to a widely accepted three-band model, and (2) to evaluate the GA-PLS spatial transferability with simulated ESA/Sentinel3/OLCI and Hyperion spectra. The GA-PLS and the three-band model yield accurate calibrations (Cal) for the SA dataset with R2above 0.98, and the corresponding validation (Val) shows relative root mean squared error (rRMSE) of less than 6.2% with narrow-band spectra. Both the GA-PLS and three-band model show stable performance for the CIN dataset (Cal: R2=0.91 and 0.77; Val: rRMSE=20.1% and 33.4%), THL dataset (Cal: R2=0.91 and 0.88; Val: rRMSE=30.1% and 33.7%), and STKR dataset (R2=0.84 and 0.82; rRMSE=29.1% and 33.2%). The results also reveal that simulated OLCI datasets degrade both the GA-PLS performance, and particularly the performance of the three-band model due to the coarser and discontinuous spectral configuration. Contrastingly, both the GA-PLS and the three-band model show improved results with the simulated Hyperion datasets. Our observation indicates that the GA-PLS model outperforms the three-band model in terms of spatial transferability; however, the three-band model has its own merits, considering its simplicity. Further analyses indicate that spectral measurement protocols, instrumentations, and inorganic suspended matter affect the GA-PLS and three-band model performances.