Recent optical remote sensing satellite missions, such as Sentinel-2 with the MultiSpectral Imager (MSI) onboard, allow the estimation of coastal water key parameters with very high spatial resolutions (down to 10 m). In this article, multiple approaches are proposed for retrieving chlorophyll-a (Chl-a) and total suspended matter (TSM) along the Adriatic and Tyrrhenian coasts in Italy, using both empirical and model-based frameworks to design regressive and neural network (NN) estimation methods. The latter proves to be more accurate on a regional scale, where standard ocean color physical models exhibit high uncertainty in their local parameterization due to the complex spectral characteristics of the observed scene. Retrieval results are encouraging for Chl-a with a coefficient of determination ${R}^{2}$ up to 0.72 with a root-mean-square error (RMSE) of 0.33 mg $\text{m}^{-3}$ , using an empirical NN. The TSM algorithms exhibit higher uncertainty, mainly due to scarcity of in situ measurements and model parameterizations, with $R^{2}= 0.52$ and RMSE = 1.95 g/m³ using NNs. The bio-optical model, used for the development of model-based algorithms, shows some inadequacies in representing the inherent and apparent optical properties for the case study areas, especially considering the different spectral features between the oligotrophic Tyrrhenian Sea and the eutrophic Adriatic Sea. This study confirms the potential of Sentinel-2 MSI products for coastal water monitoring, but it also highlights key issues to be further tackled such as the atmospheric correction impact, the need of reliable in situ measurements, and possible bathymetry effects near the shores.

中文翻译：

---

使用物理、统计和神经网络检索方法从 Sentinel-2 卫星数据进行沿海水域遥感

最近的光学遥感卫星任务，例如带有多光谱成像仪 (MSI) 的 Sentinel-2，允许以非常高的空间分辨率（低至 10 m）估计沿海水域的关键参数。在本文中，提出了多种方法来检索意大利亚得里亚海和第勒尼安海岸沿线的叶绿素 a (Chl-a) 和总悬浮物 (TSM)，同时使用经验和基于模型的框架来设计回归和神经网络 (NN) ) 估计方法。后者被证明在区域尺度上更准确，由于观测场景的复杂光谱特征，标准海洋颜色物理模型在其局部参数化中表现出高度不确定性。具有确定系数的 Chl-a 的检索结果令人鼓舞 ${R}^{2}$ 高达 0.72，均方根误差 (RMSE) 为 0.33 mg $\text{m}^{-3}$ ，使用经验神经网络。TSM 算法表现出更高的不确定性，主要是由于就地 测量和模型参数化，与 $R^{2}= 0.52$ 和 RMSE = 1.95 g/m ³使用 NN。用于开发基于模型的算法的生物光学模型在表示案例研究区域的固有和表观光学特性方面存在一些不足，特别是考虑到贫营养第勒尼安海和富营养亚得里亚海之间的不同光谱特征。这项研究证实了 Sentinel-2 MSI 产品在沿海水域监测方面的潜力，但它也强调了需要进一步解决的关键问题，例如大气校正影响、可靠就地 测量，以及靠近海岸的可能的测深效应。