**Hamid Mohebzadeh, Junho Yeom and Taesam Lee \***

Department of Civil Engineering, ERI, Gyeongsang National University, 501 Jinju-daero, Jinju, Gyeongnam 52828, Korea; hamidmohebzadeh@gnu.ac.kr (H.M.); junho.yeom@gnu.ac.kr (J.Y.)

**\*** Correspondence: tae3lee@gnu.ac.kr

Received: 20 March 2020; Accepted: 28 April 2020; Published: 30 April 2020

**Abstract:** Chlorophyll-a (Chl-a) is one of the major indicators for water quality assessment and recent developments in ocean color remote sensing have greatly improved the ability to monitor Chl-a on a global scale. The coarse spatial resolution is one of the major limitations for most ocean color sensors including Moderate Resolution Imaging Spectroradiometer (MODIS), especially in monitoring the Chl-a concentrations in coastal regions. To improve its spatial resolution, downscaling techniques have been suggested with polynomial regression models. Nevertheless, polynomial regression has some restrictions, including sensitivity to outliers and fixed mathematical forms. Therefore, the current study applied genetic programming (GP) for downscaling Chl-a. The proposed GP model in the current study was compared with multiple polynomial regression (MPR) to different degrees (2nd -, 3 rd-, and 4th-degree) to illustrate their performances for downscaling MODIS Chl-a. The obtained results indicate that GP with R<sup>2</sup> = 0.927 and RMSE = 0.1642 on the winter day and R<sup>2</sup> = 0.763 and RMSE = 0.5274 on the summer day provides higher accuracy on both winter and summer days than all the applied MPR models because the GP model can automatically produce appropriate mathematical equations without any restrictions. In addition, the GP model is the least sensitive model to the changes in the input parameters. The improved downscaling data provide better information to monitor the status of oceanic and coastal marine ecosystems that are also critical for fisheries and fishing farming.

**Keywords:** spatial downscaling; MODIS chlorophyll-a; sentinel-2A MSI; multiple polynomial regression; genetic programming

#### **1. Introduction**

Coastal marine ecosystems are the most important habitats for species that live in the world's most productive ecosystems, such as fish and marine mammals [1]. The influences of the proximity to land, large quantities of nutrients delivered via streams, and sewage discharge lead to increased susceptibility of these ecosystems to rapid changes in water quality through both anthropogenic and natural mechanisms. Therefore, it is essential to monitor water quality in coastal ecosystems to mitigate the adverse impacts of human-related activities in these environments [1,2].

A phytoplankton cell is a planktonic photosynthesizing organism [3], and phytoplankton biomass can serve an index to provide information about marine ecosystem health. Coastal ecosystems throughout the world are affected by the fast growth of the phytoplankton population, often resulting from water column stratification or increases in nutrients [2]. Harmful algal blooms, like dinoflagellate, Gymnodinium breve (commonly referred to as "red tide") produce neurotoxins such as saxitoxin and gonyautotoxin that cause water quality degradation, which have considerable consequences for marine environments such as fish death [4–7]. The chlorophyll-a (Chl-a) concentration has been recognized as a direct indicator of phytoplankton biomass because all phytoplanktons contain Chl-a and high

Chl-a concentrations show more desirable environmental conditions for phytoplankton growth [3]. Therefore, by monitoring changes in the Chl-a concentration distribution, long-term trends in the water quality of coastal and oceanic systems can be assessed to a point where the negative effects can be mitigated [2,8,9]. However, traditional techniques such as in situ field sampling, moored instruments, drifting instruments, and fluorometry used for Chl-a measurement are expensive laboratory-based instruments and have some spatiotemporal limitations.

Over the last two decades, there has been an increase in remote sensing applications as a substitute for traditional techniques for near-real-time measurements of global phytoplankton biomass, including both qualitative and quantitative estimates [10,11]. However, there are two major challenges associated with extracting information from remote sensing data: 1) the sheer amount of data, and 2) variable precision and continuity among remote sensing-derived products. [12]. To solve such issues, many techniques have been developed, such as reflectance-based classification algorithms [13], spectral band ratios [14–16], spectral band-difference algorithms [17–19], bio-optical models [20,21], and analytical techniques [22,23].

The retrieval of Chl-a concentrations in coastal areas by the abovementioned techniques is performed using coarse spatial resolution ocean color sensors such as Moderate Resolution Imaging Spectroradiometer (MODIS), Coastal Zone Color Scanner (CZCS), MEdium Resolution Imaging Spectrometer (MERIS), and Sea-viewing Wide Field-of-view Sensor (SeaWiFS). Although the high temporal resolution of these sensors (e.g., 1–2 days for MODIS, 3 days for MERIS, and 1 day for SeaWiFS) makes them suitable for continuous monitoring, their spatial resolution (e.g., 4 km for MODIS, 300 m for MERIS, and 1.1 km for SeaWiFS) is not satisfactory due to their orbital characteristics and technical configurations.

Recent studies have introduced spatial downscaling algorithms as an alternative solution to the coarse spatial resolution of ocean color sensors. Spatial downscaling has been widely used for downscaling coarse spatial resolution data by utilizing the high-resolution remote sensing reflectance measurements, for instance for land surface temperature [24–26], precipitation [27–30], and soil moisture [31,32]. Fu, et al. [33] combined coarse spatial resolution MODIS Chl-a measurements with high spatial resolution Landsat 8 OLI band combinations using a polynomial regression model (fourth-order polynomial regression) to downscale MODIS Chl-a maps from 4 km to 30 m spatial resolution. However, polynomial regression models have some restrictions, such as sensitivity to outliers and the use of fixed mathematical forms to define the relationship between the predictor and predictand variables. Machine learning (ML) approaches have been suggested to deal with these restrictions and have received increasing attention for downscaling studies as powerful alternative tools, but only limited applications for downscaling of Chl-a [34–36].

Among ML models, genetic programming (GP) have recently received much attention in a number of fields including water resource management studies [37]. The idea behind the GP was inspired by biological evolution that makes it a collection of techniques for finding the best solution in the space of possible solutions. This unique feature of GP made it a suitable technique for various water resource management applications, including ocean engineering and hydrology, hydrological forecasts, and groundwater modeling [37]. Therefore, the current study assessed the accuracy of GP for Chl-a downscaling and compared its results with the results of three multiple polynomial regression (MPR) models, including second-order (2nd), third-order (3rd), and fourth-order (4th) polynomials. The developed models were utilized for Chl-a downscaling over the western coast of South Korea.

#### **2. Study Area**

The study area was part of the Korean West Sea, which is located in the eastern part of the Yellow Sea (35◦15<sup>0</sup> <sup>−</sup> <sup>36</sup>◦30<sup>0</sup> N, <sup>125</sup>◦45<sup>0</sup> <sup>−</sup> <sup>126</sup>◦45<sup>0</sup> E; area of 10,705 km<sup>2</sup> ) (Figure 1). The Yellow Sea is a shallow marine ecosystem with the average and maximum water depths of 44 m and 103 m, respectively [38]. There is clear seasonality in sea surface temperature (SST) over the Yellow Sea, where January is the coldest month with an average SST of 4–7 ◦C and July is the warmest, with an average SST of

26–27 ◦C [39]. There are a total of 339 fish species in the Korean West Sea [40]. Over the past few years, some fish species, such as small yellow croaker, hairtail, large yellow croaker, and flatfish have exhibited continuous declines due to overharvesting, degraded marine ecosystem quality, and several unknown factors [41]. *Remote Sens.* **2020**, *12*, x FOR PEER REVIEW 4 of 28

**Figure 1.** Location of the study region in the Korean West Sea and Chl-a sampling stations.

**Figure 1.** Location of the study region in the Korean West Sea and Chl-a sampling stations. The increase in the level of the eutrophication, as a result of human activities such as dense agricultural practices along the coastal area, is another reason for environmental pollution over the study area that have significant negative effects on marine ecosystems, such as fish death, and a loss of important protein for the people dependent upon them [1]. Furthermore, in the Yellow Sea waters, there are other constituents than phytoplankton, such as inorganic particles and dissolved organic matter that are the major obstacle for simple empirical algorithms to determine the statistical relationship between Chl-a concentration and spectral bands [42]. Additionally, a very limited number of studies have attempted to investigate the optical properties of the Yellow Sea, such as phytoplankton, from the ocean The increase in the level of the eutrophication, as a result of human activities such as dense agricultural practices along the coastal area, is another reason for environmental pollution over the study area that have significant negative effects on marine ecosystems, such as fish death, and a loss of important protein for the people dependent upon them [1]. Furthermore, in the Yellow Sea waters, there are other constituents than phytoplankton, such as inorganic particles and dissolved organic matter that are the major obstacle for simple empirical algorithms to determine the statistical relationship between Chl-a concentration and spectral bands [42]. Additionally, a very limited number of studies have attempted to investigate the optical properties of the Yellow Sea, such as phytoplankton, from the ocean color images. As a result, monitoring the Chl-a concentration, as an important index to evaluate the extent of eutrophication, at fine resolution is a crucial task in this area.

#### **3. Materials and Methods**

The main objective of this research was to develop an approach for MODIS Chl-a downscaling and produce Chl-a concentration maps for complex coastal regions. Figure 2 displays the detailed explanation of the procedure used. The downscaling approach was accomplished in four steps: (1) remote sensing data, including MODIS Chl-a at 4 km (defined as Y4k) and S-2A at 10 m (defined as X10), were acquired, and S-2A data were resampled to 4 km MODIS resolution (denoted as X4k); (2) the most important S-2A band combinations (X4k) were chosen by utilizing the support vector machine recursive feature elimination (SVM-RFE) method; (3) MODIS Chl-a downscaling from 4 km to 10 m was performed by regressing X4k to Y4k, calculating the residual at 4 km (ε4k), and adding the interpolated

residual (ε10) to the estimated fine-resolution Chl-a (Yˆ <sup>10</sup>); (4) the obtained downscaled maps were compared with visual comparison, validated with in situ data, and all the applied methods were assessed using sensitivity analysis. A complete explanation of the aforementioned steps is described in Section 3.1, Section 3.2, Section 3.3, Section 3.4. *Remote Sens.* **2020**, *12*, x FOR PEER REVIEW 6 of 28

**Figure 2.** Downscaling workflow. Note that the goal of the present research is to downscale Moderate Resolution Imaging Spectroradiometer (MODIS) Chl-a from coarse resolution (4 km) to high resolution (10 m).
