Deriving VIIRS High-Spatial Resolution Water Property Data Over Coastal and Inland Waters Using Deep Convolutional Neural Network

: The Visible Infrared Imaging Radiometer Suite (VIIRS) onboard the Suomi National Polar-orbiting Partnership (SNPP) satellite has been a reliable source of ocean color data products, including ﬁve moderate (M) bands and one imagery (I) band normalized water-leaving radiance spectra nL w ( λ ). The spatial resolutions of the M-band and I-band nL w ( λ ) are 750 m and 375 m, respectively. With the technique of convolutional neural network (CNN), the M-band nL w ( λ ) imagery can be super-resolved from 750 m to 375 m spatial resolution by leveraging the high spatial resolution features of I1-band nL w ( λ ) data. However, it is also important to enhance the spatial resolution of VIIRS-derived chlorophyll-a (Chl-a) concentration and the water diffuse attenuation coefﬁcient at the wavelength of 490 nm ( K d (490)), as well as other biological and biogeochemical products. In this study, we describe our effort to derive high-resolution K d (490) and Chl-a data based on super-resolved nL w ( λ ) images at the VIIRS ﬁve M-bands. To improve the network performance over extremely turbid coastal oceans and inland waters, the networks are retrained with a training dataset including ocean color data from the Bohai Sea, Baltic Sea, and La Plata River Estuary, covering water types from clear open oceans to moderately turbid and highly turbid waters. The evaluation results show that the super-resolved K d (490) image is much sharper than the original one, and has more detailed ﬁne spatial structures. A similar enhancement of ﬁner structures is also found in the super-resolved Chl-a images. Chl-a ﬁlaments are much sharper and thinner in the super-resolved image, and some of the very ﬁne spatial features that are not shown in the original images appear in the super-resolved Chl-a imageries. The networks are also applied to four other coastal and inland water regions. The results show that super-resolution occurs mainly on pixels of Chl-a and K d (490) features, especially on the feature edges and locations with a large spatial gradient. The biases between the original M-band images and super-resolved high-resolution images are small for both Chl-a and K d (490) in moderately to extremely turbid coastal oceans and inland waters, indicating that the super-resolution process does not change the mean values of the original images. is desirable to also super-resolve VIIRS Chl-a and K d (490) to 375 m resolution, particularly over turbid coastal oceans and inland waters. In this study, we further develop and evaluate high-spatial resolution Chl-a and K d (490) products based on super-resolved nL w ( λ ) images of the VIIRS ﬁve M-bands. Speciﬁcally, high-spatial resolution Chl-a and K d (490) are derived from super-resolved nL w ( λ ) spectra using the same Chl-a and K d (490) algorithms as implemented in the MSL12 ocean color data processing system High-spatial resolution Chl-a and K d (490) products are compared and evaluated with the original data of 750 m resolution. Sea (Granule V2019105051208), and La Plata River Estuary (Granule V2020058172016), and the super-resolved 750-m-resolution data are compared pixel-by-pixel with the original nL w ( λ ) images for validation. The results show that the mean, median, and standard deviation (STD) of the super-resolved/original nL w ( λ ) ratio in the three regions on average


Introduction
Ocean color environmental data records (EDRs, or Level-2 data) from the Visible Infrared Imaging Radiometer Suite (VIIRS) onboard the Suomi National Polar-orbiting Partnership (SNPP) satellite provide high quality ocean color remote sensing data source for science research and ocean environment monitoring. The normalized water-leaving radiance spectra nL w (λ) [1][2][3] of the five moderate (M) bands (M1-M5 at wavelengths of 410, 443, 486, 551, and 671 nm) with a spatial resolution of 750 m are the fundamental ocean color products derived from VIIRS. In addition, VIIRS also provides high-spatial resolution visible imagery (I) bands with 375 m resolution, and especially nL w (λ) of the Remote Sens. 2021, 13, 1944 3 of 22 In the previous study, Liu and Wang (2021) [16] trained and evaluated neural networks for super-resolving VIIRS M-bands nL w (λ) using ocean color data from the Bohai Sea and Baltic Sea separately, and found that the networks trained from these two regions are equivalent. However, waters over neither the Bohai Sea nor Baltic Sea are extremely turbid. Therefore, the networks trained from these two regions lack proper skills to sharpen the ocean color nL w (λ) images over highly turbid waters. Turbid coastal oceans and inland waters usually contain a large amount of sediment, which has significant effects on the optical property of the water column. Turbid waters in the world oceans are usually located near the world's major river estuaries [33,34], for example, Yangtze River Estuary, La Plata River Estuary, Amazon River Estuary, Meghna River Estuary, and so on. In the coastal region, the water turbidity is highly dynamic and closely associated with the atmosphere, ocean, and land variability, such as cyclones, algae blooms, and flood-driven river plumes. High-resolution Chl-a and K d (490) images in the coastal region can help to understand water column property and benthic processes such as primary productivity, coral reef ecosystem, nutrient dynamics, and river dynamics. In this study, the data from the La Plata River Estuary are added into the training dataset, so that the networks are more robust to deal with highly turbid waters. Therefore, the objectives of this study are twofold, i.e., first, to retrain the network with the dataset of the La Plata River Estuary; and second, to develop and evaluate the high-spatial resolution Chl-a and K d (490) with newly trained networks. The rest of the paper is organized as follows. The deep CNN models, Chl-a and K d (490) algorithms, as well as datasets used for training, evaluation, and testing, are described in Section 2. The results of network training, evaluation, and testing for Chl-a and K d (490) products are presented in Section 3. Finally, discussions and conclusion are provided in Section 4.

VIIRS Ocean Color EDR Data
We use the VIIRS-SNPP ocean color EDR (or Level-2) data processed by the NOAA Ocean Color Team (https://www.star.nesdis.noaa.gov/socd/mecb/color/, accessed on 14 May 2021) using the MSL12 ocean color data processing system [13,35]. MSL12 is the NOAA's official ocean color data processing system for VIIRS, and it has been used to process satellite ocean optical, biological, and biogeochemical data using various atmospheric correction algorithms [35]. The atmospheric correction algorithms include the near-infrared (NIR) approach [1,2,36], the shortwave infrared (SWIR) algorithm [3], and the NIR-SWIR combined algorithm [37]. In particular, the NIR-SWIR-based atmospheric correction in MSL12 has been shown to significantly improve satellite ocean color data over global open oceans and turbid coastal/inland waters [38][39][40][41]. In this study, the NIR-SWIR combined atmospheric correction algorithm [37] is used to derive VIIRS-SNPP ocean color nL w (λ) spectra.
As an important result of the atmospheric correction in the ocean color data processing [1][2][3], normalized water-leaving radiance nL w (λ) spectra [2,[42][43][44] measure the outgoing radiances backscattered from oceans/waters at different wavelengths. VIIRS-SNPP ocean color product data include nL w (λ) spectra at the seven M-bands (750 m) at wavelengths of 410, 443, 486, 551, 671, 745, and 862 nm, and one I-band (375 m) nL w (λ) at 638 nm. It was reported that the high resolution nL w (638) can reveal more detailed fine features of water optical, biological, and biogeochemical properties than those with M-band nL w (λ) [4]. In fact, the red I-band nL w (638) data can be used in various ocean color applications to study fine spatial resolution dynamic features, particularly over turbid coastal oceans and inland waters [16,45,46].
VIIRS ocean color products also include Chl-a and K d (490), which are frequently used by users to monitor ocean environment, e.g., ocean/water biological productivity [47,48] and harmful algal bloom (HAB) monitoring [49,50]. VIIRS Chl-a and K d (490) data are derived from nL w (λ) spectra [13,35]. Specifically, the K d (490) algorithm is a hybrid of clear ocean (standard) model and turbid ocean model, so that the K d (490) product for both clear and turbid ocean waters can be accurately retrieved [9]. The VIIRS Chl-a algorithm uses the ocean color index (OCI) method, which has been proven to be more stable and accurate for clear (low Chl-a) oceans [6,7]. In this study, these two K d (490) and Chl-a algorithms are implemented in the super-resolution application (Python) to generate 375 m resolution products, so that the super-resolved K d (490) and Chl-a can be compared with original data.
The I1 band nL w (638) data have strong signals in coastal oceans and inland waters, but generally have very weak signals in open oceans. Therefore, in this study, we use ocean color data derived from VIIRS-SNPP in coastal oceans and inland waters, including the Bohai Sea, Baltic Sea, La Plata River Estuary, Great Lakes, Chesapeake Bay, and Gulf of Mexico. A single VIIRS-SNPP ocean color EDR file (granule) comprises an~85 seconds of orbit segment spanning 48 scans. Each individual scan contains 32 and 16 rows for the I-band and M-band, respectively, with one row for each detector. Individual rows contain a total of 6400 and 3200 samples in the I-band and M-band data, respectively. Consequently, each VIIRS ocean color EDR granule has a total of 6400 × 1536 and 3200 × 768 image elements in the I-band and M-band data, respectively, which cover a ground swath of approximately 3060 km wide.

Convolution Neural Networks and Training
Because the spatial features in VIIRS I1-band nL w (638) images are highly correlated with those in M-bands, Liu and Wang (2021) [16] employed the deep CNN to obtain the high-frequency content from nL w (638), and use it to super-resolve M-band nL w (λ) imageries. In their study, VIIRS ocean color data over the Bohai Sea and Baltic Sea were used to train the networks for super-resolution of coastal ocean color nL w (λ) images. However, a significant discrepancy is found between the original M-band nL w (λ) images and super-resolved images when the networks are applied to extremely turbid ocean areas. The poor performance of the network over extremely turbid waters is attributed to the limitation of the training dataset: it only contains ocean color data from the Bohai Sea and Baltic Sea, which are not highly turbid waters. Therefore, the networks lack the proper skills to super-resolve images of extremely turbid oceans.
In this study, VIIRS ocean color data from the extremely turbid La Plata River Estuary are included in the training dataset. We retrain the same CNN models [51], and the network architecture is shown in Figure 1, and is briefly described as follows. The M-band nL w (λ) spectra are first bilinearly upsampled to the same dimension as the I1 band, nL w (638). The upsampled M-band nL w (λ) are then concatenated with nL w (638) along the horizontal axis, and serve as the input of the network. It is the concatenation step that adds the high-spatial resolution I-band information to the procedure. The input data go into the first convolution layer, a rectified linear unit (ReLU), iteration of residual blocks, the last convolution layer, and an additive skip connection at the end. For the convolution layers, we use 128 feature maps and 3 × 3 convolution kernels, as using a small kernel size can benefit from weight sharing and reduction in computational costs [23]. In addition, zero-padding is used so that the convolved image has the same dimension as the input image. The ReLU function simply truncates all negative values to zeroes in the output. The module of residual block employs the popular ResNet architecture [52], which uses skip connections to make it faster to train very deep networks.
Sens. 2021, 13, x FOR PEER REVIEW As in Liu and Wang (2021) [16], we retrain five net M-band nLw(λ). Network training requires a large amo but we do not have 375 m resolution nLw(λ) data as gro make an assumption that super-resolving nLw(λ) image (original scale) is self-similar to super-resolving images scale [16,51]. With this assumption, we train the network scale, i.e., VIIRS M-band nLw(λ) images are synthetically d m, and used as coarse-resolution inputs, and the original truth. The networks trained with synthetic data have be resolve nLw(λ) images on the original scale [16]. Forty-fi EDR data (Table 1) are used as the training dataset, and patches of 32 × 32 pixels. We use 90% of patches for trai As in Liu and Wang (2021) [16], we retrain five networks with one network for each M-band nL w (λ). Network training requires a large amount of data as a training dataset, but we do not have 375 m resolution nL w (λ) data as ground truth. Therefore, we need to make an assumption that super-resolving nL w (λ) images from 750 m to 375 m resolution (original scale) is self-similar to super-resolving images on a degraded 1500 m to 750 m scale [16,51]. With this assumption, we train the networks on a degraded 1500 m to 750 m scale, i.e., VIIRS M-band nL w (λ) images are synthetically downsampled from 750 m to 1500 m, and used as coarse-resolution inputs, and the original images are treated as the ground truth. The networks trained with synthetic data have been proven to work well to super-resolve nL w (λ) images on the original scale [16]. Forty-five granules of VIIRS ocean color EDR data (Table 1) are used as the training dataset, and they are preprocessed into small patches of 32 × 32 pixels. We use 90% of patches for training the model weights and 10% for validation. The network and training are implemented in the Keras frameworks (https://github.com/keras-team/keras, accessed on 14 May 2021) with TensorFlow as backend.

Training Networks
The performance of the deep CNN relies heavily on the training datasets. VIIRS ocean color data over the Bohai Sea and Baltic Sea were used to train the networks for superresolution of coastal ocean color images [16]. The maximum K d (490) values of the Bohai Sea and Baltic Sea are usually~2.0 m −1 and~4.0 m −1 , respectively. However, in extremely turbid coastal and inland lake regions, for example, the La Plata River Estuary [53], K d (490) values can reach up to~6.0-7.0 m −1 [54]. Therefore, super-resolution networks trained with the Bohai Sea and Baltic Sea data could not be able to accurately super-resolve nL w (λ) images over extremely turbid waters. We have to re-train the CNNs by also including nL w (λ) spectra data from highly turbid water regions such as La Plata River Estuary in the training datasets.
Forty-five granules (Table 1) of VIIRS-SNPP ocean color data from the Bohai Sea, Baltic Sea, and La Plata River Estuary were used to re-train five networks (one for nL w (λ) at each M-band), and they are referred to as networks with new training. In comparison, the networks trained previously with only the Bohai Sea and Baltic Sea data [16] are referred to as old training. The difference between the new training and old training is only in the training dataset, while the CNN model and the settings are kept the same. It is noted that the ocean color satellite images contain lots of missing data owing to clouds, high sun glint, high sensor-zenith angle, and so on. In our treatment of the missing data in the process of CNN, we assign the mean of all valid pixels in the images to all missing data points. Figure 2 shows the comparison of the original images (750 m resolution) with superresolved images (375 m resolution) of both old training and new training for nL w (443), nL w (551), and nL w (671), as well as K d (490) and Chl-a on 27 February 2020, in the La Plata River Estuary. It can be seen that, for nL w (443) and nL w (671), the original images are essentially the same as the super-revolved images with both old training and new training. However, for nL w (551), the radiances with old training (Figure 2g) are obviously underestimated in the La Plata River Estuary (color looks more orange), while the data from new training (Figure 2l) are significantly improved and look very similar to the original image. As the training dataset of old training contains only data from the Bohai Sea and Baltic Sea, the networks of old training do not have the skills to super-resolve the images of highly turbid water regions. Consequently, Chl-a data are underestimated with old training (the color looks more greenish in the La Plata River Estuary in Figure 2j), as Chl-a are derived from nL w (443), nL w (486), and nL w (551) [7]. The training dataset of new training includes the data from the extremely turbid La Plata River Estuary, so Chl-a are significantly improved (Figure 2o). On the other hand, K d (490) data are more correlated to nL w (671) [9], so the differences between using old training and new training are generally small (Figure 2i,n). ever, for nLw(551), the radiances with old training (Figure 2g) are obviously underestimated in the La Plata River Estuary (color looks more orange), while the data from new training ( Figure 2l) are significantly improved and look very similar to the original image. As the training dataset of old training contains only data from the Bohai Sea and Baltic Sea, the networks of old training do not have the skills to super-resolve the images of highly turbid water regions. Consequently, Chl-a data are underestimated with old training (the color looks more greenish in the La Plata River Estuary in Figure 2j), as Chl-a are derived from nLw(443), nLw(486), and nLw(551) [7]. The training dataset of new training includes the data from the extremely turbid La Plata River Estuary, so Chl-a are significantly improved ( Figure 2o). On the other hand, Kd(490) data are more correlated to nLw(671) [9], so the differences between using old training and new training are generally small (Figure 2i and n).  To quantitatively evaluate the difference between the results of using old training and new training, Figure 3 shows the line plots of nL w (443), nL w (551), nL w (671), K d (490), and Chla along line A (pink line) marked in Figure (Figure 3b), which indicate that the region is indeed extremely turbid in the La Plata River Estuary [54]. In addition, nL w (551) values are also abnormally high (~5.0 mW cm −2 µm −1 sr −1 ), and they are higher than those in the Bohai Sea and Baltic Sea (~4.0 mW cm −2 µm −1 sr −1 ) [16]. Therefore, the networks trained in the Bohai Sea and Baltic Sea (old training) underestimate nL w (551) by~7% in the La Plata River Estuary. Figure 3c and d shows the percent difference, i.e., 100 × (super-resolved − original)/original, of Chl-a and K d (490) with old training (blue cross) and new training (red circle) along line A noted in Figure 2a. As a result, Chl-a concentrations are underestimated by~21% with old training (Figure 3c), while K d (490) are only underestimated by~1% with old training (Figure 3d). With new training, both K d (490) and Chl-a values are significantly improved with essentially zero biases. Therefore, it is concluded that including the data from highly turbid water regions in the training dataset (i.e., covering all possible data ranges) is necessary for the network models to accurately super-resolve satellite-measured K d (490) and Chl-a images in various coastal and inland water environments.

Network Validation
As described in the previous section, the network training assumes that super-resolv ing nLw(λ) images from 750 m to 375 m resolution are self-similar to super-resolving im ages on a degraded 1500 m to 750 m scale. This assumption requires that the nLw(λ) rela tionships between bands of different spatial resolutions are self-similar within the rele vant scale range, i.e., the transfer of spatial details in nLw(λ) from high-spatial resolution to low-spatial resolution bands does not depend on the spatial resolution itself, but de pends only on the relative resolution difference. In fact, the assumption is supported by the literature on self-similarity in image analysis [55,56], and has been proven to be effec tive in image super-resolution applications [16,51]. Following the same validation proce dure by Liu and Wang (2021) [16], quantitative evaluations of the networks of new training are performed by super-resolving nLw(λ) images on a degraded 1500 m to 750 m scale and the original 750 m spatial resolution nLw(λ) data are treated as ground truth to evalu ate the super-resolved nLw(λ) images. The networks obtained from new training are ap plied on down-sampled 1500 m resolution images in the Baltic Sea (Granule V2015226105214), Bohai Sea (Granule V2019105051208), and La Plata River Estuary (Gran ule V2020058172016), and the super-resolved 750-m-resolution data are compared pixel by-pixel with the original nLw(λ) images for validation. The results show that the mean median, and standard deviation (STD) of the super-resolved/original nLw(λ) ratio in the three regions on average are 0.995, 0.997, and 0.056, respectively. Therefore, the perfor mance is consistent with the validation results tested previously in the Baltic Sea and Bo

Network Validation
As described in the previous section, the network training assumes that superresolving nL w (λ) images from 750 m to 375 m resolution are self-similar to super-resolving images on a degraded 1500 m to 750 m scale. This assumption requires that the nL w (λ) relationships between bands of different spatial resolutions are self-similar within the relevant scale range, i.e., the transfer of spatial details in nL w (λ) from high-spatial resolution to low-spatial resolution bands does not depend on the spatial resolution itself, but depends only on the relative resolution difference. In fact, the assumption is supported by the literature on self-similarity in image analysis [55,56], and has been proven to be effective in image super-resolution applications [16,51]. Following the same validation procedure by Liu and Wang (2021) [16], quantitative evaluations of the networks of new training are performed by super-resolving nL w (λ) images on a degraded 1500 m to 750 m scale, and the original 750 m spatial resolution nL w (λ) data are treated as ground truth to evaluate the super-resolved nL w (λ) images. The networks obtained from new training are applied on down-sampled 1500 m resolution images in the Baltic Sea (Granule V2015226105214), Bohai Sea (Granule V2019105051208), and La Plata River Estuary (Granule V2020058172016), and the super-resolved 750-m-resolution data are compared pixel-by-pixel with the original nL w (λ) images for validation. The results show that the mean, median, and standard deviation (STD) of the super-resolved/original nL w (λ) ratio in the three regions on average are 0.995, 0.997, and 0.056, respectively. Therefore, the performance is consistent with the validation results tested previously in the Baltic Sea and Bohai Sea [16].
As the goal of this study is to super-resolve K d (490) and Chl-a images, we also evaluate the performance of the networks on K d (490) and Chl-a data on a degraded 1500 m to 750 m scale. Specifically, super-resolved 750 m nL w (λ) data generated above from the down-sampled 1500 m resolution are used to calculate K d (490) and Chl-a data (superresolved). Pixel-by-pixel comparisons between the super-resolved and original K d (490) and Chl-a data of the same resolution are evaluated. The results show that the mean, median, and STD of the super-resolved/original K d (490) ratio in the three regions on average are 1.003, 1.001, and 0.077, respectively. For Chl-a, the mean, median, and STD of the super-resolved/original ratio in the three regions on average are 0.995, 0.998, and 0.096, respectively. For both parameters, the biases between the super-resolved and original data are quite small, but their STD values are larger than those in nL w (λ). In addition, the performance of super-resolved K d (490) and Chl-a in the La Plata River Estuary shows no significant difference than that in the Baltic Sea and Bohai Sea.

Super-Resolved K d (490) and Chl-a in the Baltic Sea
The networks with new training are used to super-resolve to VIIRS M-band nL w (λ) images at the Baltic Sea from the spatial resolution of 750 m to 375 m. As the Baltic Sea is not a highly turbid region, the difference between using new training and old training can be neglected. The super-resolved nL w (λ) images of the VIIRS five M-bands in the Baltic Sea have been evaluated in the previous study [16]. In this study, we mainly focus on high-resolution K d (490) and Chl-a products, which are derived from the super-resolved nL w (λ) spectra. Figure Figure 4a,b, which shows that the super-resolution process does not change the mean value of those features.
Similarly, fine spatial structures of Chl-a are also enhanced in the super-resolved images (Figure 4c,d). Because the Chl-a algorithm uses nL w (λ) at the blue and green bands, which is less correlated to nL w (638) [16], the pattern of the Chl-a features looks different from those of K d (490) (as expected for the two different products). The high-concentration Chl-a filaments (yellow and red) are much sharper and thinner in the super-resolved image (Figure 4d), and some of the finer features that are not available in the original Chl-a image now appear in the super-resolved Chl-a image. As shown in K d (490) images, the super-resolving process does not change the mean value of these Chl-a features, so the color of these Chl-a features is generally the same in the super-resolved and original images.   Figure 4a for further analysis.
The enhancement of the super-resolved K d (490) images is further evaluated quantitatively at the pixel level. Figure 5 shows the line plots of the super-resolved and original K d (490) images along line B (pink line in Figure 4a). The original and super-resolved K d (490) values are compared to nL w (638) along the line to demonstrate the performance of the network. The differences between K d (490) and nL w (638) are usually quite large in magnitude in the Baltic Sea. However, we mostly care about the K d (490) changes along line B for the purpose of performance evaluation. Therefore, we normalized nL w (638) and K d (490) values along line B to the value of the first point on the left (by taking the ratio), so that we can easily compare their variations along the line. Figure 5a shows the line plots of the original and super-resolved normalized K d (490) values along line B, as well as the normalized nL w (638) values on the same line (scale noted in the right). It can be seen that the super-resolved (normalized) K d (490) images (red line) have more high-frequency variations than the original (normalized) K d (490) images (blue line). These high-frequency variations correspond to the fine spatial features in the super-resolved K d (490) images. It is also noted that the high-frequency (normalized) K d (490) variations are mostly consistent with those in nL w (638) (green line). In fact, the high correlation coefficient of 0.9661 between super-resolved K d (490) and nL w (638) indicates that the high-frequency variations in K d (490) have been successfully transferred from the super-resolved M-band, which essentially learned from the I1 band (nL w (638)).
Remote Sens. 2021, 13, x FOR PEER REVIEW 11 of Kd(490) have been successfully transferred from the super-resolved M-band, which esse tially learned from the I1 band (nLw(638)).  Furthermore, we quantitatively compare the original K d (490) and Chl-a images with the super-resolved images in the Baltic Sea. As each pixel in the original image corresponds to four pixels in the super-resolved image, the density-scatter plots of the super-resolved images versus original images by corresponding pixels for K d (490) and Chl-a are shown in Figure 6, and they are both very close to the 1:1 line. The statistics of the ratio between the original K d (490) (or Chl-a) images and the corresponding super-resolved images are also calculated. The mean ratios of the super-resolved/original K d (490) and Chl-a image are 0.997 and 1.008, respectively, indicating that the biases between the original K d (490) (or Chl-a) images and the corresponding super-resolved images are very small (<~1%). The STD values of the super-resolved/original image ratios for K d (490) and Chl-a are 0.045 and 0.062, respectively. In fact, these variations are mostly caused by the added high-resolution features in the super-resolved K d (490) and Chl-a images.

Tests in Other Coastal and Inland Waters
The performance of the networks is also tested in four coastal and inland wate gions, i.e., the Bohai Sea, Chesapeake Bay, Lake Erie, and Gulf of Mexico, which inc both sediment-dominate and phytoplankton-dominate waters. The networks with training are applied to VIIRS ocean color data in the four regions to super-resolve nL images of five M-bands, and then high-resolution Kd(490) and Chl-a are derived from super-resolved nLw(λ) spectra for evaluations.

The Bohai Sea Region
The Bohai Sea is a major marginal sea of approximately 78,000 km 2 on the east c of China. In the Bohai Sea, the ocean dynamics and water properties are dominated marily by river runoff, semidiurnal/diurnal tides, and seasonal monsoons, and so on As the Yellow River flows into the Bohai Sea, both sediments and phytoplankton blo have significant effects on the water optical property in this region [58]. Shi et al. (2 [59] also reported that the spring-neap tide can make significant Kd(490) variations in Bohai Sea, as the tidal current plays an important role on sediment resuspension. In a tion, the Kd(490) features can also be enhanced by the sand ridges on the bottom o eastern Bohai Sea [58]. Figure 7 shows Kd(490) and Chl-a images in the Bohai Sea

Tests in Other Coastal and Inland Waters
The performance of the networks is also tested in four coastal and inland water regions, i.e., the Bohai Sea, Chesapeake Bay, Lake Erie, and Gulf of Mexico, which include both sediment-dominate and phytoplankton-dominate waters. The networks with new training are applied to VIIRS ocean color data in the four regions to super-resolve nL w (λ) images of five M-bands, and then high-resolution K d (490) and Chl-a are derived from the super-resolved nL w (λ) spectra for evaluations.

The Bohai Sea Region
The Bohai Sea is a major marginal sea of approximately 78,000 km 2 on the east coast of China. In the Bohai Sea, the ocean dynamics and water properties are dominated primarily by river runoff, semidiurnal/diurnal tides, and seasonal monsoons, and so on [57]. As the Yellow River flows into the Bohai Sea, both sediments and phytoplankton blooms have significant effects on the water optical property in this region [58].   [59] also reported that the spring-neap tide can make significant K d (490) variations in the Bohai Sea, as the tidal current plays an important role on sediment resuspension. In addition, the K d (490) features can also be enhanced by the sand ridges on the bottom of the eastern Bohai Sea [58]. Figure 7 shows K d (490) and Chl-a images in the Bohai Sea on 9 September 2018, as well as the ratio of the super-resolved/original image. K d (490) data have a large dynamic range in the region, as its maximum value can reach as high as~4 m −1 (Figure 7a). With the strong influence of the tidal current and wind waves [59], there are lots of fine dynamic features in satellite ocean color images. For example, the finger-shaped features on the bottom of the K d (490) image are induced by the sand ridges on the ocean floor [60]. These features are significantly enhanced by the super-resolved K d (490) image. To compare the original K d (490) images with the super-resolved images, we calculate the pixel-by-pixel super-resolved/original K d (490) ratio (Figure 7b). It can be seen in most parts of the image where K d (490) are uniform, and ratios of super-resolved/original are close to 1.0 (green color). In the locations with some significant K d (490) features, ratios of superresolved/original deviate from 1.0. For example, the edges of the finger-shaped feature induced by sand ridges [60] are clearly modified, so that the feature edges are sharpened in the super-resolved image. The mean and STD of the super-resolved/original K d (490) ratio for the region are 1.003 and 0.056, respectively (Table 2), which indicate that there are little biases between the super-resolved and original image. The Chl-a image of the Bohai Sea looks quite uniform (Figure 7c). Consequently, there is less enhancement in the super-resolved Chl-a image (Figure 7d). The mean and STD of the super-resolved/original Chl-a ratio for the region are 0.999 and 0.060, respectively. of super-resolved/original deviate from 1.0. For example, the edges of the finger-shaped feature induced by sand ridges [60] are clearly modified, so that the feature edges are sharpened in the super-resolved image. The mean and STD of the super-resolved/origina Kd(490) ratio for the region are 1.003 and 0.056, respectively (Table 2), which indicate tha there are little biases between the super-resolved and original image. The Chl-a image o the Bohai Sea looks quite uniform (Figure 7c). Consequently, there is less enhancement in the super-resolved Chl-a image (Figure 7d). The mean and STD of the super-re solved/original Chl-a ratio for the region are 0.999 and 0.060, respectively.

The Chesapeake Bay Region
The Chesapeake Bay is the largest estuary in the United States, and one of the most productive water bodies in the world. There are about 150 rivers, creeks, and streams flowing into the Chesapeake Bay, and the top five largest rivers, Susquehanna, Potomac, Rappahannock, York, and James rivers, contribute~90% of discharge. As a large amount of freshwater (~2300 m 3 s −1 on average) is discharged into the Chesapeake Bay, the dissolved and particulate materials can have a significant effect on the water property [61], and strongly influences phytoplankton production in the bay [48]. In addition, tidal forcing also plays an important role in sediment, biological, and geological processes in the Chesapeake Bay [62]. The upper Chesapeake Bay is extremely turbid because of the estuarine turbidity maximum (ETM) of the Susquehanna River [9,63]. Figure 8 shows K d (490) and Chl-a images acquired on 3 March 2018, Granule V2018062175339, as well as the ratio of the super-resolved/original image in the Chesapeake Bay. It is noted that the K d (490) value is so high (~5 m −1 ) in the upper Chesapeake Bay (Figure 8a) that, with old training, nL w (551) data are usually underestimated by~4%, and Chl-a are generally underestimated by~9%. The networks with new training work quite well; the super-resolved/original ratios for K d (490) (Figure 8b) and Chl-a (Figure 8d) are both close to 1.0 (green color) in the upper Chesapeake Bay. Figure 8a also shows lots of K d (490) features in the main stem of the Chesapeake Bay as well as in the tributaries. In general, it is much more turbid in the tributaries than in the main stem, as shown in Figure 8a. In the middle Chesapeake Bay, K d (490) has the minimum in the center of the main stem, and gradually increases to the east and west shorelines. The super-resolution of all these water turbidity features is very important for routine monitoring of the Chesapeake Bay environment. It can be seen in Figure 8b that the super-resolution process is performed on these features with a significant K d (490) gradient all around the Chesapeake Bay and the tributaries. On the other hand, Chl-a are generally uniform except in the middle bay region where the phytoplankton bloom is found on the west side of the Chesapeake Bay (Figure 8c). The super-resolution of the Chl-a is mainly performed in the middle bay (Figure 8d). In fact, as shown in Table 2, the mean values of the super-resolved/original ratios in K d (490) and Chl-a are 1.010 and 1.004, respectively.
cant Kd(490) gradient all around the Chesapeake Bay and the tributaries. On the other hand, Chl-a are generally uniform except in the middle bay region where the phytoplankton bloom is found on the west side of the Chesapeake Bay (Figure 8c). The super-resolution of the Chl-a is mainly performed in the middle bay (Figure 8d). In fact, as shown in Table 2, the mean values of the super-resolved/original ratios in Kd(490) and Chl-a are 1.010 and 1.004, respectively.

Lake Erie
Lake Erie, the southern most of the Great Lakes, is generally not extremely turbid (K d (490) <~3 m −1 ), except for the west side of the lake [64,65]. However, there are often significant water quality issues owing to rapid growth of algae, which is mainly caused by runoff of fertilizer and manure spread on large farm fields [66][67][68]. These problems not only threaten the lake's ecosystem, but also the economic activities of the surrounding areas. Indeed, Lake Erie provides a primary drinking water source, and supports many recreational, tourism, and fishery activities. In addition, although most algae blooms in Lake Erie are harmless, the blue-green algae Microcystis produce toxins, and have harmful effects on human, birds, marine mammals, fish, and local economies [68,69]. Figure 9 shows K d (490) and Chl-a images of Lake Erie on 29 April 2018 (Granule V2018119182603). There are some turbid features (K d (490) >~2.0 m −1 ) mainly on the west side of the lake, but for most area of the lake, waters are non-turbid (or relatively clear, K d (490) <~1.0 m −1 ) (Figure 9a). Figure 9b shows the super-resolved/original ratio of K d (490) in Lake Erie. It can be seen that differences of the super-resolved and original K d (490) images are small for most pixels, except for the feature edges. The Chl-a image looks quite uniform (Figure 9c can be seen that differences of the super-resolved and original Kd(490) images are small for most pixels, except for the feature edges. The Chl-a image looks quite uniform ( Figure  9c), and there are less Chl-a feature enhancements than those in the super-resolution of Kd(490) image (Figure 9d). The mean values of the super-resolved/original ratio in Kd(490) and Chl-a are 0.998 and 0.995 (Table 2), indicating that the super-resolution process does not change the general values of the original Kd(490) and Chl-a images.

The Gulf of Mexico Region
As the fourth longest river in the world (~6300 km), the Mississippi River discharges at a speed of 18,400 m 3 s −1 on average [70], and carries ~230 million tons of sediment into the Gulf of Mexico annually [71]. In the meantime, the Mississippi River brings a significant amount of nutrients into the Gulf of Mexico each year. As river plumes have significant effects on coastal marine ecosystems, including nutrient dynamics, biological productivity, water quality, and pollution hazards [72], high-resolution ocean color imageries can significantly benefit routine monitoring and understanding of river plume dynamics.

The Gulf of Mexico Region
As the fourth longest river in the world (~6300 km), the Mississippi River discharges at a speed of 18,400 m 3 s −1 on average [70], and carries~230 million tons of sediment into the Gulf of Mexico annually [71]. In the meantime, the Mississippi River brings a significant amount of nutrients into the Gulf of Mexico each year. As river plumes have significant effects on coastal marine ecosystems, including nutrient dynamics, biological productivity, water quality, and pollution hazards [72], high-resolution ocean color imageries can significantly benefit routine monitoring and understanding of river plume dynamics. Figure 10 shows the K d (490) and Chl-a images of 19 November 2019 (Granule V2019323185429), which cover the turbid Mississippi River plume. As shown in Figure 10a, waters are usually highly turbid (K d (490) >~2.5 m −1 ) around the Mississippi River Delta, but the water is clear in the open ocean. There are some finer features on the boundary between the turbid and clear waters. Figure 10b also shows that there are significant super-resolution enhancements on the turbid-clear water boundaries, and on some coastal features as well. The Chl-a image provides similar features to those of K d (490) around the Mississippi River Delta (Figure 10c), and the enhancement in the super-resolved image mainly occurs on the edge of these features. The mean values of the super-resolved/original ratios in K d (490) and Chl-a are 1.002 and 1.015, respectively, in the Gulf of Mexico region ( Table 2). ary between the turbid and clear waters. Figure 10b also shows that there are significant super-resolution enhancements on the turbid-clear water boundaries, and on some coastal features as well. The Chl-a image provides similar features to those of Kd(490) around the Mississippi River Delta (Figure 10c), and the enhancement in the super-resolved image mainly occurs on the edge of these features. The mean values of the superresolved/original ratios in Kd(490) and Chl-a are 1.002 and 1.015, respectively, in the Gulf of Mexico region (Table 2). In summary, from the results of the previous four examples, both Kd(490) and Chl-a images in the turbid coastal oceans and inland waters show many enhanced fine spatial features. While Chl-a images look relatively more uniform than Kd(490), significant superresolution enhancements occur in both images. This is particularly useful for turbid coastal oceans and inland waters. In addition, super-resolution occurs mainly on pixels with significant Kd(490) and Chl-a spatial features, especially on the feature edges and locations with a large gradient. Overall, however, there are few differences on average between the original (750 m) Kd(490) and Chl-a images and the super-resolved (375 m) images in moderately and highly turbid coastal oceans and inland waters, showing that the super-resolution process does not change the mean values of the original ocean color product images. In summary, from the results of the previous four examples, both K d (490) and Chl-a images in the turbid coastal oceans and inland waters show many enhanced fine spatial features. While Chl-a images look relatively more uniform than K d (490), significant superresolution enhancements occur in both images. This is particularly useful for turbid coastal oceans and inland waters. In addition, super-resolution occurs mainly on pixels with significant K d (490) and Chl-a spatial features, especially on the feature edges and locations with a large gradient. Overall, however, there are few differences on average between the original (750 m) K d (490) and Chl-a images and the super-resolved (375 m) images in moderately and highly turbid coastal oceans and inland waters, showing that the super-resolution process does not change the mean values of the original ocean color product images.

Discussions and Conclusions
As satellite images from more sensors have become available, there are now many research activities dedicated to super-resolving remote sensing images. Liu and Wang (2021) [16] used the deep CNN to sharpen VIIRS M-band nL w (λ) images using the highspatial resolution features from I1-band nL w (638) data. While nL w (λ) spectra of the five M-bands are primary parameters directly derived from VIIRS measurements, ocean/water biological and biogeochemical products Chl-a and K d (490) are important water property data and have been widely used in the science and user communities for various research and applications [47][48][49][50]. In this study, we further develop and evaluate high-resolution Chl-a and K d (490) products based on super-resolved VIIRS nL w (λ) images of the five Mbands. The deep CNN method is retrained with the training dataset from the Bohai Sea, Baltic Sea, and La Plata River Estuary as well, so that the performance of networks is significantly improved over extremely turbid coastal oceans and inland waters. In fact, the new trained networks, which cover wide water optical types, are generally applicable to various highly turbid coastal oceans and inland waters (other than the region with the training dataset). Indeed, the evaluation results show that a super-resolved K d (490) image is much sharper than the original one, and has more detailed fine spatial structures. Although Chl-a features look different from K d (490) (as expected), the fine spatial features are also enhanced in the super-resolved Chl-a images. The high K d (490) and Chl-a filaments are much sharper and thinner in the super-resolved images, and some of the finer features that are not available in the original (750 m) images now appear in the super-resolved (375 m) images. The networks are also tested in four other coastal and inland waters. The evaluation results show that the super-resolution occurs mainly on pixels of K d (490) and Chl-a spatial features (i.e., spatial variations), especially on the feature edges and locations with a large spatial gradient. It is also concluded that the super-resolution process does not change the mean value of the original images, while adding high-frequency features in the super-resolved water property product images.
It should be noted that the high-resolution K d (490) and Chl-a images are derived from super-resolved nL w (λ) spectra images. As analyzed in Liu and Wang (2021) [16], the high-spatial resolution I1 band, nL w (638), is highly correlated to nL w (λ) of the VIIRS five M-bands, so that the networks are trained to extract high-frequency variations from nL w (λ) at the I1 band and transfer to those at M-bands. However, both K d (490) and Chl-a algorithms use nL w (λ) at multi-spectral bands, and their relationships to that at the I1 band are not linear [7,9]. As can be seen in Figure 4, K d (490) and Chl-a images usually show different patterns owing to differences in the algorithms, which indicate that they are not directly correlated. Thus, directly super-resolving K d (490) and Chl-a images from coarse resolution to fine resolution is not as effective as that for nL w (λ). In fact, the evaluation results show that super-resolved nL w (λ) at the VIIRS five M-bands are quite accurate [16] and, consequently, VIIRS-derived K d (490) and Chl-a images from super-resolved nL w (λ) spectra perform better than those from the direct super-resolution approach for K d (490) and Chl-a.
It is also noted that the networks have the limitation when they are applied to super-resolve ocean color images in clear open ocean. Owing to the sensor characteristics, nL w (638) data generally have very weak signals with noticeable noise in the clear open ocean. Consequently, the networks add the data noise in the super-resolved nL w (λ), K d (490), and Chl-a images. For example, in the open ocean outside of the Chesapeake Bay ( Figure 8) and in the Gulf of Mexico (Figure 10), the value of K d (490) is~0.1 m −1 , and the super-resolved K d (490) and Chl-a images show some noise on the lower-right corner.
Because the performance of the deep CNN relies on the training dataset, it is recommended to verify and validate the super-resolved images when the networks are applied to other coastal and inland water regions, although the results show that the networks are also applicable to various coastal and inland water regions (different from the training dataset). Coastal and inland waters around the world generally show different optical properties owing to various suspended particulate matter types (e.g., sediment, phytoplankton, colored dissolved organic matter, and so on) as well as concentrations, which lead to complex M-band nL w (λ) spectra in VIIRS ocean color images. Abnormal nL w (λ) outside the range of the training dataset could result in inaccurate super-resolution in nL w (λ), K d (490), and Chl-a images. The water turbidity seems to be an important factor that affects the super-resolution process. In our studies, the validation results indeed show that the networks with old training underestimate nL w (551) and Chl-a in the La Plata River Estuary and the upper Chesapeake Bay region. By including the La Plata River Estuary data in the training dataset (new training), it not only resolves the discrepancies in the La Plata River Estuary, but also in the highly turbid upper Chesapeake Bay region. Therefore, it is important to include various water types in the training dataset. We emphasize that it is not necessary to include all coastal/inland water data for the training, rather it is important to evaluate and validate the super-resolved images when the networks are applied to a new region of interest.
In the case of a discrepancy found between the super-resolved and original images when applied to a new region, network models need to be re-trained by including data of the region (or water type) in the training dataset. Fortunately, re-training the networks does not mean restarting from scratch. Instead, the networks can be trained incrementally, i.e., starting from the previously trained CNN weights, and re-training with only the new data. With this approach, the computation time to train with new dataset is significantly reduced, while the networks still retain all previously learned skills. By adding more and more data to the training dataset, the network model weights are gradually fine-tuned, and the network eventually will have the skills to super-resolve ocean color images of global coastal oceans and inland waters. Furthermore, we can also train the networks to utilize nL w (λ) spectra with high resolution images from many other satellite sensors. For example, OLCI on the Sentinel-3A/3B has spatial resolution of 300 m, and the Operational Land Imager (OLI) on Landsat-8 has spatial resolution of 30 m. In fact, the networks have the capability to simultaneously super-resolve high resolution ocean color data from multiple sources of different satellite sensors. Therefore, the high-spatial resolution imageries from multi-sensors and multi-bands can be fused into networks to conduct the super-resolving ocean color data processing.