Accurate estimation of coastal and in-land water quality parameters is important for managing water resources and meeting the demand of sustainable development goals. The water quality monitoring based on discrete water sample analysis is limited to specific locations and becomes less effective to offer a synoptic view of the water quality variability at different spatial and temporal scales. The optical remote sensing techniques have proved their ability to provide a comprehensive and synoptic view of water quality parameters. In conjugation with other products, the optical remote sensing data products can be utilized for the effective management of water bodies while addressing the socio-economic issues faced by local governments and states. In recent years, multiple machine-learning (ML) models have been reported on the estimation of water quality using remote sensing data, but their performance is limited when extended to diverse water types within coastal and inland water environments. In this study, we present an ensemble machine-learning model for estimating the primary water quality parameters in coastal and inland waters, such as Chlorophyll-a (Chl-a) concentration, colored dissolved organic matter (aCDOM440), and Turbidity. It utilizes the in-situ measurements to train and optimize the ensemble machine-learning models for the spectral measurements data (400–700 nm) provided by MODIS-Aqua, Sentinel-2 Multi Spectral Instrument (MSI), and PlanetScope (Planet). To develop the prediction models, these in-situ measurements data were split into two parts: a training dataset (70 %) and a testing dataset (30 %). The ensemble machine-learning models were validated using the 5-fold cross-validation method. These models were trained and tested against distinct datasets encompassing a broad range of variations in water quality parameters collected from open ocean, coastal and inland waters. The validation results demonstrated a superior performance of the present ensemble ML models compared to other ML models (Chl-a: R2 = 0.96, RMSE = 4.93, MAE = 2.89; aCDOM440: R2 = 0.93, RMSE = 0.057, MAE = 0.025; Turbidity: R2 = 0.95, RMSE = 4.52, MAE = 1.009). To realize the importance of this study, the ensemble ML models were applied to MODIS-Aqua monthly composite measurements from 2003 to 2022 and captured pronounced seasonal variations in water quality parameters (WQP) and Water Quality Index (WQI). For instance, in the Gulf of Khambhat, turbidity decreased at an annual average rate of ∼0.08 NTU and Chl-a increased at an annual average rate of ∼0.004 mg m−3 for the past 20 years. Furthermore, we investigated the occurrences of Noctiluca scintillans (here after N. scintillans) bloom between 2019 and 2021 near the fin fish cage culture sites in Mandapam, on the southeast coast of Tamil Nadu, within the Gulf of Mannar, India which serves as a documentation of the Harmful Algal Bloom (HAB) incidents. The performance of ensemble model is further demonstrated using Planet images from inland turbid waters of the Muthupet lagoon (Brackish water) and Adyar river (Urban River) and MSI image from Chilika lagoon. The proposed ensemble ML models proved as an effective method for accurately estimating the WQP and WQI products and capturing their spatial and temporal variations in regional and global waters, which forms an important tool for sustainable development and management of coastal and inland water environments.

中文翻译：

---

沿海和内陆水域的长期水质评估：一种使用卫星数据的集成机器学习方法


准确估计沿海和内陆水质参数对于管理水资源和满足可持续发展目标的需求非常重要。基于离散水样分析的水质监测仅限于特定位置，并且在提供不同空间和时间尺度上水质变化的概要视图效果较差。光学遥感技术已经证明它们能够提供水质参数的全面和概要视图。与其他产品结合使用，光学遥感数据产品可用于有效管理水体，同时解决地方政府和州面临的社会经济问题。近年来，已经报道了多种机器学习 （ML） 模型使用遥感数据估算水质，但当扩展到沿海和内陆水域环境中的不同水类型时，它们的性能受到限制。在这项研究中，我们提出了一个集成机器学习模型，用于估计沿海和内陆水域的主要水质参数，例如叶绿素-a （Chl-a） 浓度、有色溶解有机物 （aCDOM440） 和浊度。它利用原位测量来训练和优化集成机器学习模型，以用于 MODIS-Aqua、Sentinel-2 多光谱仪器 （MSI） 和 PlanetScope （Planet） 提供的光谱测量数据 （400–700 nm）。为了开发预测模型，这些原位测量数据分为两部分：训练数据集 （70%） 和测试数据集 （30%）。集成机器学习模型使用 5 倍交叉验证方法进行验证。 这些模型在不同的数据集上进行了训练和测试，这些数据集涵盖了从公海、沿海和内陆水域收集的水质参数的广泛变化。验证结果表明，与其他 ML 模型相比，当前集成 ML 模型的性能更优越 （Chl-a： R2 = 0.96， RMSE = 4.93， MAE = 2.89;aCDOM440： R2 = 0.93， RMSE = 0.057， MAE = 0.025;浊度：R2 = 0.95，RMSE = 4.52，MAE = 1.009）。为了认识到这项研究的重要性，将集成 ML 模型应用于 2003 年至 2022 年的 MODIS-Aqua 月度综合测量，并捕获了水质参数 （WQP） 和水质指数 （WQI） 的明显季节性变化。例如，在过去 20 年中，Khambhat 湾的浊度以 ∼0.08 NTU 的年平均速率下降，而 Chl-a 以 ∼0.004 mg m-3 的年平均速率增加。此外，我们还调查了 2019 年至 2021 年间在泰米尔纳德邦东南海岸曼达帕姆的有鳍鱼网箱养殖点附近发生的夜蛾藻水华（这里在 N. scintillans之后）水华的情况，位于印度马纳尔湾内，作为有害藻华 （HAB） 事件的文件。使用 Muthupet 泻湖（咸水）和 Adyar 河（城市河）内陆浑浊水域的 Planet 图像以及 Chilika 泻湖的 MSI 图像进一步证明了集成模型的性能。所提出的集成 ML 模型被证明是准确估计 WQP 和 WQI 产品并捕获它们在区域和全球水域的时空变化的有效方法，这构成了沿海和内陆水域环境可持续发展和管理的重要工具。