**3.2 Retrieval algorithms for Chl-a concentrations through Sentinel-2 data**

In the studies conducted by Marzano et al. [76], Sentinel-2 data played a crucial role in the detection of water quality. Their study was focused on Case-II waters, as per Morel and Prieur water classification [77]. In coastal areas, indeed, water quality is mainly conditioned by Chl-a and TSM concentrations variations and their study was focused on different retrieval approaches to these quantities.

**Figure 3.** *In situ observation points locations of ARPAs' dataset.*

*Coastal Water Quality: Hydrometeorological Impact of River Overflow and High-resolution… DOI: http://dx.doi.org/10.5772/intechopen.104524*

The region of interest in this study is central-northern Italy over Tyrrhenian and the Adriatic Sea. The *in situ* observations of Chl-a concentrations were provided by Italian regional environmental agencies named ARPA (Agenzia Regionale per la Protezione Ambientale), for the regions of Tuscany, Lazio, Abruzzo, and Veneto, and covered the time period between 2016-08-04 and 2018-04-19. In **Figure 3**, the spatial distribution of the dataset along the Italian coasts is shown.

They analysed the use of both empirical and model-based regressive algorithms to retrieve IOPs. More in detail:


The atmospheric correction software used is ACOLITE, with the Dark Spectrum Fitting (DSF) enabled. **Figure 4** reports an example of RGB composite images for Top Of Atmosphere and Bottom Of Atmosphere reflectance before and after atmospheric correction, respectively.

Focusing on the Empirical Regressive algorithm (EmpReg), developed in [76], the retrieval of Chl-a concentrations was defined through the following:

$$\mathcal{F}\_{MBR} = \frac{\max\left(R\_{wlB1}, R\_{wlB2}\right)}{R\_{wlB3}},\tag{1}$$

where the numerator is the maximum between B1 and B2 (blue bands) waterleaving reflectance and the denominator is the water-leaving reflectance for B3 (green band). Indeed, the bands more sensitive to chlorophyll presence were

**Figure 4.**

*RGB Sentinel-2 remote-sensing reflectance images over the Adriatic coast in the Marche region before (a) and after (b) the atmospheric correction using the ACOLITE software.*

considered to retrieve Chl-a concentrations. This blue-to-green reflectance maximum band ratio (MBR) model is among the most used ones in literature [78, 79].

The empirical regressive retrieval algorithm, which is the optimal regressive formula found with respect to the area of analysis and dataset used in the paper, is defined as:

$$\bar{C}\_{C\&\mu} = a\_1 \exp\left(-a\_2 r\_{\text{MRR}}\right) \tag{2}$$

where *a*1=59.795 mg/m3 and *a*2 = 4.559.

In the following **Figure 5**, is reported the scatterplot of chlorophyll-a (Chl-a) *in situ* measurements (mg m−3) with respect to Sentinel-2 MSI water-leaving blue-togreen maximum band ratio (MBR) in the Tyrrhenian and the Adriatic Sea. **Figure 5** highlights the non-linearity that characterises the relationship between Chl-a concentrations and MBR.

However, statistical regression algorithms show limitations in handling nonlinearity and non-monotonicity, and to overcome these limitations, Marzano et al. [76] used also neural networks. This allowed the exploitation of data contained in several MSI spectral channels of Sentinel-2 products (from B1 to B8A) and spatiotemporal information. In the same way as the previously described methods, also in this case, two neural network-based algorithms were tested:


**Figure 5.** *Measured MBR with respect to in situ Chl-a concentration for the whole training dataset.*

The experiments conducted in the work lead to observing better results for NN algorithms when trained with empirical data, rather than with synthetic ones. Although a test to be performed on a wider dataset would be needed. The results obtained through the empirical regressive algorithms with MBR, instead, did not always provide an accurate estimation of the Chl-a concentration, depending on the higher turbidity of Tyrrhenian coastal waters. This was probably related to turbidity conditions of water, which can impact the effectiveness of estimation.
