1. Introduction
Covering around 71% of the Earth’s surface, oceans play a major role in the global climate system, [
1]. The study of sea surface temperature and salinity is important to understand how oceans communicate with land and atmosphere, but also for the understanding of marine ecosystems and weather prediction, [
2]. Sea surface salinity (SSS) and temperature (SST) are also relevant in the study of estuarine processes (mixing of fresh and sea water), stratification, hypoxia, organic matter, or algal blooms, among others, [
3]. The SSS and SST data collection has typically been done by means of static buoys, drifters, and ship-based systems, [
4]. The problems that these measurements present are related to poor extended data coverage. Moreover, the estimation of SST and SSS near the coast, where the detail needed might be higher due to the development of different near-shore processes and human activities, is difficult.
Different satellite missions have focused over the past decades on the measurement of oceanographic characteristics to overcome the issues that the in situ measuring techniques present. Satellites provide worldwide coverage of ocean and land phenomena, which is particularly relevant for the extraction of time series and general trends, but also for the study of localised events. One of these satellite operations is the European Space Agency’s (ESA) SMOS (Soil Moisture Ocean Salinity) mission, which provides measurements of both soil moisture and ocean salinity by using a Microwave Imaging Radiometer with Aperture Synthesis (MIRAS), [
5]. The spatial resolution is 35 km at the centre of the field of view. On the other hand, NASA’s Aquarius mission focuses also on sea surface salinity from space, [
6]. Its spatial resolution is 150 km.
The MODIS–Aqua mission (Moderate Resolution Imaging Spectroradiometer), [
7], views the complete Earth’s surface every 2 days, obtaining information in 36 spectral bands. It started operations in 2000, and its role is relevant in the development of Earth system models. Bands capture information useful for the study of aerosols boundaries and properties, ocean colour, phytoplankton and biochemistry, water vapour, temperature, and much more. Data is acquired in three different spatial resolutions: 250 m, 500 m and 1000 m, having the bands used for ocean reflectance the latter. MODIS data has been used for the extraction of SST datasets.
Other sensors, such as SeaWiFS (Sea-Viewing Wide Field-of-View Sensor) [
8], have been used for the extraction of SSS data from their colour information. It was designed to collect ocean biological data, mainly chlorophyll, and was active from 1997 to 2010. It recorded information in eight bands and its resolution ranges from 1100 to 4500 m. Moreover, the NOAA CoastWatch mission, [
9], derived SST products from different instruments in a set of satellites: Visible Infrared Imaging Radiometer Suite (VIIRS) in the Joint Polar Satellite System, Advanced Very High Resolution Radiometer (AVHRR) in the MetOp, or Advanced Baseline Imager (ABI) from GOES-16, among others. In every case, the highest available resolution is in the order of 1000 m. The low available resolution in built-for-purpose ocean observation satellites is clear from the information provided above. This influences the use of remote data for coastal and detailed oceanographic applications, where more detail might be needed. The different satellite missions and principal characteristics detailed in this introduction are summarised in
Table 1.
SST is normally estimated from satellite measurements following different retrieval algorithms, depending on the type of sensor. For example, [
10] summarises the SST retrieval algorithm for Aqua. This is based in the correction of several factors affecting temperature at the surface, leading to an accurate estimation of SST through clouds. The corrections included in the algorithm are due to atmosphere, wind, incident angle, sea ice, sun glitter, land contamination, salinity, and ground interferences. The inclusion of corrections leads to higher SST-derived accuracy. The spatial resolution of the satellites considered in [
10] is about 50 km. Moreover, [
11] evaluates the accuracy of SST at different frequencies. It is found that radiometer noise and geophysical errors are less representative in lower frequencies and are independent of latitude.
Although some satellite data provide users with direct SST information and retrieval algorithms, researchers are developing techniques to further estimate SST information and make more accurate predictions. The use of neural networks to estimate oceanographic information is a good example of further efforts in predicting SST. [
12] presents an artificial neural network model to predict SST based on measurements of SST in the previous day. A 2-year dataset is predicted with errors below 0.5 °C in most cases. [
13] predicts SST in the western Mediterranean Sea using in situ data to train the neural network: sea level pressure, wind components, air temperature, dew point temperature, and total cloud cover. A total of 45 years of data are provided, and these are combined with satellite-derived SST data from Pathfinder and MODIS at a resolution of 4 km. The neural network provides accurate predictions of seasonal and interannual SST variability, and the methodology is then used to reconstruct incomplete SST satellite images.
The method proposed in [
14] combines numerical estimations with SST measurements to produce real-time SST in the Indian Ocean using a wavelet neural networks. NOAA’s AVHRR SST derived data is used in this study together with other remotely obtained information, linked to numerical data provided by the Indian National Centre for Ocean Information Services (INCOIS). The results provide accurate daily, weekly, and monthly SST values. Finally, some innovative techniques have compared the SST obtained by using AVHRR information and with data collected by surfers, [
15]. However, the resolution of the AVHRR is around 1 km. While this is useful for oceanographic applications offshore, more detail is needed near the coast for specific applications.
Looking now into sea surface salinity estimation, the amount of literature in this area is less extensive than that for SST. In the last decades, different research projects have focused on the extraction and validation of SSS from the information provided by satellites, and have combined this with the use of machine learning techniques. [
16] investigates the ability of two algorithms (least squared and second polynomial order) to retrieve SSS from MODIS satellite data. The information is compared with in situ data from the South China Sea. They found out that the least square algorithm can be used to retrieve SSS from MODIS data. In line with this, other studies have given a step further, and implemented neural networks to predict SSS. [
17] uses 9000 salinity records from four research vessels matched with MODIS–Aqua satellite data to train a neural network that estimates SSS in the mid-Atlantic area, obtaining relevant values for estuarine areas.
In a further step, [
3] developed a neural network to estimate SSS in the Gulf of Mexico. This study uses ocean reflectance data reprocessed by NASA from MODIS–Aqua and SeaWiFS data, and the resolution is in the range of 1000 m. The satellite information includes atmospheric correction, and the time window allowed to match in situ and satellite data is ±6 h. Other studies present interesting approaches to solving the salinity characterisation issue: [
18] presents a 6-year study on the salt distribution in the Western Mediterranean Sea from SMOS information, reducing errors from previous studies. The study is based in the use of debiased non-Bayesian retrieval Data Interpolating Empirical Orthogonal Functions (DINEOF), and multifractal fusion. Both the North Atlantic and Mediterranean Sea are mapped using this technique, providing reduced error in SMOS SSS. The different methodologies described in this section are summarised in
Table 2.
Reference [
19] recently presented an article on the extraction of ocean salinity from space. The paper presents the difference in development of the satellite-observed salinity missions and developments compared with sea surface temperature, which has been done for the past four decades. The paper states the requirement of using temperature in order to estimate salinity, as temperature impacts the brightness temperature of the ocean observed from space. In contrast, the research presented in the following sections shows how salinity can be estimated without the need of including temperature as input.
For the purpose of this study, generic optical satellite information has been chosen: ESA’s Sentinel-2 Level 1-C. As pointed out previously, there are built-for-purpose satellite missions which focus on the observation of ocean colour, such as Sentinel-3, [
20], or NASA’s MODIS–Aqua, [
7]. However, Sentinel-2 allows to produce higher resolution SSS and SST datasets from its 10–60 m band resolutions (compared with the 300 m from Sentinel-3, and 1000 m for MODIS–Aqua). Moreover, we include band and general image information, in order for the network to learn atmospheric relationships, such as cloud coverage, water vapour, or mean angle of incidence. The in situ information has been obtained from the Copernicus Marine Environment Monitoring Service (CMEMS), which has been operational since May 2015. This is part of the Copernicus programme, which is an initiative from the European Union for the establishment of a European capacity for Earth Observation and Monitoring, [
4]. Copernicus is composed of three components: Space, Insitu, and Services. The first one includes the European Space Agency’s (ESA) Sentinels, as well as other contributing missions operated by national or international organisations. The Insitu component collects information from different monitoring networks around Europe, such as weather stations, ocean buoys, or maps. This information is key to calibrate and validate satellite data.
Taking into account the research done in the area, in the present paper, we aim at obtaining SST and SSS from TOA data, thus developing a methodology that reduces the amount of pre-processing needed from the satellite data, such as atmospheric corrections. Moreover, the use of optical satellite data provides a wide variety of options where the proposed methodology is applicable. Special attention is paid to the use of raw data, allowing wider use and faster extraction of information. Moreover, the time window between in situ measurement and equivalent satellite image has been reduced in this paper to 1 h, which is considerably lower than that used in previous research papers. The reduced time window linked to the high resolution provided by Sentinel-2, provides higher reliability than previous methodologies. Moreover, this research allows the prediction of sea surface salinity independently of sea surface temperature, based on the hypothesis that the neural network is built with enough information to obtain key salinity relationships with visual parameters.
The paper is composed of three main sections, and is structured as follows:
Section 2 presents the methodology followed in the paper, including extraction of in situ and satellite data, matching process, data usage and coverage, and neural network details.
Section 3 includes the discussion of prediction results for sea surface temperature and salinity in interpolation and extrapolation problems, as well as a sensitivity analysis, and neural network evaluation over a region of interest. Finally,
Section 4 presents the work conclusions.
4. Conclusions
This paper presents a methodology to obtain high-resolution values of sea surface salinity and temperature in the global ocean by using satellite data without any band data pre-processing or atmospheric corrections. Sentinel-2 Level 1-C Top of Atmosphere reflectance data has been used here. A deep neural network has been built to link band information with in situ data from different buoys, vessels, drifters, and other platforms around the world. This in situ information has been obtained from the Copernicus Marine In Situ platform.
The neural network presented in this paper outperformed classical architectures tested for regression problems. The Sentinel-2 data was processed using Google Earth Engine in areas of 100 m m around relevant platforms. Accurate salinity values are estimated for the first time without using temperature as input in the network. Salinity results depend only on direct satellite observations. However, a clear dependency on temperature ranges is observed, with less accurate estimations for location where ocean temperature falls below 10 C. The reasons for this is unclear from the findings presented in this paper, but could be linked to the amount of available data in areas with lower temperatures and the ability of the network to predict them, or the presence of other physical phenomena such as freezing and mixing. In order to overcome this issue more information is needed in higher latitudes and areas where other phenomena, such as mixing processes and freezing, are representative. Extended information would also help reducing the amount of uncertainties on outliers in prediction.
The neural network has good interpolation and extrapolation capabilities. Mean absolute errors of 0.8 PSU and 1.3 C with correlation coefficients of and for salinity and temperature, respectively, are obtained in test locations. The most common error for both temperature and salinity is 0.4 C and 0.4 PSU. A sensitivity analysis was also performed, showing that outliers are present in areas where the number of observations is lower. The network is tested in a complete tile over the Guadalquivir River mouth, in the south-west coast of Spain. Results show clear seasonal patters, as well as a sensible interaction between river discharge and ocean intrusions that matches records from local buoys. Issues with sun glitter are observed in some images, although the general behaviour remains. This issue could be sorted by using SWIR information to identify sun glitter in each image. This methodology is relevant for detailed coastal and oceanographic applications. The time for data pre-processing is reduced significantly compared with previous approaches. Moreover, the methodology is applicable to a wide range of satellites, as the information is directly obtained from Top of Atmosphere information.