Coastal upwelling is a process whereby cool and deep nutrient-rich waters are brought to the ocean surface, primarily as a result of wind driven Ekman transport (e.g., Kämpf and Chapman, 2016). The nutrients brought to the surface by upwelling drive increased primary productivity and in turn support higher trophic levels including fish (e.g., Cushing, 1971). These coastal upwelling regions are among the most highly productive marine ecosystems regions and rich fishing grounds around the globe (Cushing, 1971; Barber, 2001). In the Western Indian Ocean (WIO), coastal upwelling systems are seasonal (only present for part of the year), driven by changing wind directions over the year, leading to changes in their productivity (e.g., Kämpf and Chapman, 2016b). They play a key role in regulating regional ecosystem productivity and sustaining food security and livelihoods for millions of people via fishing activity (Bakun et al., 1998; Jacobs et al., 2020a; Jacobs et al., 2020b; Jebri et al., 2020; Jebri et al., 2022a).

The Somali Coastal Upwelling (SCU) is considered to be the largest upwelling system in the WIO (Chatterjee et al., 2019) and the fifth most important coastal upwelling in the world ocean (DeCastro et al., 2016). The Somali upwelling occurs during the Southwest Monsoon season (from May to September) as the strong Findlater jet (low-level atmospheric jet) wind blows southwesterly (Findlater, 1971) and the positive alongshore wind stress causes offshore Ekman transport (Schott et al., 2009; Varela et al., 2015). The Findlater jet and the southwest monsoon winds drive the Somali Current (SC) northward (reversing its direction from the northeast monsoon season) (Schott and McCreary, 2001; Chatterjee et al., 2019). Part of the northward flowing SC separates from the coast at around 3-4°N and flow eastward to form a clockwise gyre called the “Southern Gyre” (Mccreary et al., 1996; Chatterjee et al., 2019). A second part of the SC continues farther north before deviating to the east at around 10°N where it interacts with the “Great Whirl”, a strong and large anticyclonic gyre (Beal and Donohue, 2013; Lakshmi et al., 2020).

The SCU results in upwelled cold subsurface water and significant biological productivity with increased nutrient concentration and enhanced chlorophyll-a (Chl, a proxy of phytoplankton biomass) levels (Mccreary et al., 1996; Wiggert et al., 2005; Lakshmi et al., 2020). The phytoplankton bloom surface signature of this upwelling region is generally wedge shaped, due to the offshore deviation of the SC and the interaction with the Great Whirl and Southern Gyre (Mccreary et al., 1996; Beal and Donohue, 2013). However, the SC, its wedges and associated gyres spread the areal extent of this upwelled productive waters over a wider region offshore (Baars et al., 1998; Lakshmi et al., 2020).

Although most indicators of upwelling activity (e.g., enhanced Chl, and reduced Sea Surface Temperature [SST]) can be observed in a synoptic way using satellite observations, the areal extent of the SCU productivity remains difficult to delineate exactly from month to month or year to year with the human eye. This task can also be highly time consuming. Machine learning (ML) approaches have proven to be efficient for automatic detection of spatial features and facial recognition (e.g., Chen et al., 2016; Zhang et al., 2018; Cheng et al., 2019). They have been successfully used in oceanographic applications such as plankton image classification (e.g., Zheng et al., 2017; Pastore et al., 2020), currents and SST clustering (e.g., Richardson et al., 2003; Liu and Weisberg, 2005), and assessing the links between current dynamics and productivity (Jebri et al., 2022b). ML approaches fall in to two broad categories: supervised learning, where the desired outputs of the training dataset are known and unsupervised learning (e.g., Bengio et al., 2013), where they are not. Spatial delineation problems can be addressed by using unsupervised ML clustering methods which classify a population as a number of subsets (or clusters). Clustering works on the principle of ensuring that data points within an assigned group are more similar to each other than those in the other groups. Examples of ML clustering approaches include DBSCAN (Ester et al., 1996) and K-means (Macqueen, 1967). Spatial delineation based on ML clustering techniques were successfully applied to identify regions with distinct marine biological activity from satellite ocean colour data (Ardyna et al., 2017). Other traditional thresholding approaches have also been used in the past for spatial delineation problems such as defining marine biogeochemical regions (Devred et al., 2007; Reygondeau et al., 2013). However, ML based clustering has the advantage of higher discriminant power (i.e., distinguishing between the different clusters) as compared to other traditional methods (Jouini et al., 2016).

Due to the importance of fishing for economic stability and food security (Taylor et al., 2019) as well as its dependence on the dynamic seasonal upwelling (e.g., Bakun et al., 1998), a potential service is proposed called “Upwelling Watch” to provide updates/alerts to interested parties concerning the areal extent and extremes of upwelling productivity as identified with ML clustering. An initial proposal for the methodology is laid out in this manuscript. In this study, an ML clustering technique based on K-means, a space-partitioning algorithm, is applied to daily remote sensing data off the Somali coast in order to identify the areal extent and classify extremes of upwelling productivity.

In this study, two K-means based clustering approaches for upwelling detection were used, leading to slightly different clustering performance. One K-means clustering model was fitted also using information for SLA (i.e., Chl/SST/SLA), the other was fitted without this information (i.e., Chl/SST only). A spatial comparison of both clustering models is included in Figure 8 , using an example from one day of NRT data, also showing the underlying variables used by the clustering technique. A reasonable correlation in the spatial patterns of the three variables over the upwelling region is seen, however the inclusion of SLA in the K-means model leads to slightly worse performance. This could relate to two factors; the SLA model picks up more features that do not directly relate to the targeted coastal upwelling (e.g., SC associated gyres) and secondly the coarser spatial resolution (25 km vs 4-5 km for the other two variables). This leads to, when a region boundary is more dominantly SLA defined, spatial boundaries that are not smooth and instead take the coarser resolution of the SLA product as well as the omission of smaller scale features. It may be possible in future that an improved resolution SLA product (e.g., from the upcoming SWOT mission) could be a useful addition to this upwelling identification system; alternatively, SLA data from models could be another possible source of this information.

Three parameters associated with upwelling (Chl, SST, and SLA) were directly explored in this manuscript. Current and wind vectors were also explored in a preliminary analysis although they were not found to have a strong correlation with the Chl and SST variables. Other variables have also been shown to impact Chl productivity in upwelling regions. For example, precipitation has been shown to have an indirect impact on Chl productivity through influence on nutrient input via riverine discharge into coastal regions (e.g., Shafeeque et al., 2019). Precipitation data was not incorporated into our methodology as precipitation primarily affects the coastal region (the majority of our study region is open ocean) and is currently of limited spatial resolution (e.g., 1° for the Global Precipitation Climatology Product, Huffman et al., 2001). Aerosols and dust have also been shown to have some very limited impact on chl productivity in the Somali offshore region (Shafeeque et al., 2017), and so they are not incorporated into our methodology.

There are a small number of limitations that affect the remote sensing data used here, although they should have limited impact. Remotely sensed Chl-a cannot be retrieved when there is cloud cover, as clouds block visible light, this is partially compensated by the use of an L4 dataset. The L4 GlobColour dataset fills missing data (due to cloud coverage) by using the most recently available value in a grid cell. SST data is similarly affected by cloud cover, at infrared wavelengths, although this is similarly compensated by use of an L4 dataset that incorporates data from microwave sensors. Additionally, Chl-a values are typically overestimated near the coast due to the impact of atmospheric aerosols and dust as well as additional constituents in the water column, including Coloured Dissolved Organic Matter (CDOM) (Schollaert et al., 2003; Hyde et al., 2007; Mélin et al., 2007). Regional studies have shown that radiometric biases can exceed ±5% for data within 25km of the coast (Bulgarelli et al., 2017). SLA data is also known to be impacted near the coast, where altimeters can be contaminated by land falling within the footprint of the altimetry instruments as well by the impact of tides and other high frequency events; data within 15km of the coast are generally considered inaccurate (The Climate Change Initiative Coastal Sea Level Team, 2020). These limitations close to the coast are largely mitigated by the study area primarily being open ocean and by the use of outlier removal.

The application of the K-means model fitted using historical reprocessed data to operational data performs well and is within the bounds produced by the historical reprocessed data (both seasonally and interannually, Figures 6 , 7 ). This indicates the model fit is likely to continue to be applicable to incoming reprocessed data for the foreseeable future, provided there are no major changes in the primary Chl/SST relationships in this region. It is also feasible that this method could be adapted in future to detect such changes. However, assessing the overall performance of these K-means based models can be somewhat challenging in two ways. First, although there is good theoretical understanding of upwelling drivers and relevant biogeochemical response, in situ data, which compared to satellite data have the advantage of sampling the ocean subsurface, are rarely available in this region. Conversely, while ocean model data is available at depth, models are currently poor at describing vertical (upwelling) velocities, so there is a dependency on secondary parameters (e.g., nutrient and Chl concentrations indicative of the biological response). These secondary parameters can have a spatial coverage different from the actual upwelling site. Comparison in future with fishing data may allow some assessment of the performance, assuming a higher abundance of fish in these productive upwelling regions, although the limited spatial resolution of available fishing datasets (e.g., up to 0.5°; Pauly et al., 2020) would make this challenging at present. The second factor challenging performance assessment is the use of traditional clustering metrics (e.g., Calinski-Harabasz score here); as the data itself changes temporally, the metrics will also reflect this variability rather than solely clustering performance (e.g., Wu et al., 2009). The data considered in each “scene” (i.e., one day of data) differs between days. This will happen on a day-to-day basis, seasonally, and also interannually, as the variables change in response to drivers beyond simply upwelling. This will in turn lead to “apparent” changes in clustering performance, in traditional clustering metrics (e.g., the Calinski-Harabasz score), that do not necessarily reflect any change in true performance. New metrics, or normalisation of existing metrics, that are robust to this may help improve assessment of such clustering approaches in future (e.g., Wu et al., 2009).

In a potential “Upwelling Watch” service, being able to not only identify upwelling but also classify extreme events will be of great use to potential end users. For the example found in Figure 9 part of the upwelling region is identified as a high extreme in Chl and low extreme in SST. In order to classify upwelling events as extremes, a number of thresholds were explored, in this case based on quantiles at 5-15% and 75-95% of the data ( Figure 7 ). Each of these thresholds thus corresponds to a different proportion of data classified as being extreme. However, the optimum proportion has not been identified in this study. Instead, it may vary depending on the subject of interest, for example for potential fisheries end users it may depend on the type of fish being targeted, noting that there can be lags between Chl productivity and fish abundance (Menon et al., 2019; Kizenga et al., 2021). As such in a potential “Upwelling Watch” service, the thresholds for this may be a user selectable option. Alternatively, to allow a more informed comparison and selection of these threshold options, future work on comparing these thresholds with fishing data may help identify the ideal thresholds.

The clustering method is based on identification of regions within Chl/SST space. As such, other ocean features such as mesoscale eddies and oceanic fronts may have a similar signature to upwelling. Evidence of the detection of an eddy can be seen in Supplementary Video 1 between the 17^th and 21^st of August, classified by the method as part of the upwelling surround or pre/post upwelling region. The analysis of fronts in this region is the focus of ongoing work although the main thermal fronts in this region have been shown to correspond with the typical Somali upwelling signature (e.g., Wang et al., 2021). When applying this method, care should be taken to avoid contamination of the signal by eddies and fronts; if it is necessary to separate these features in future work, joint application of automated eddy and front detection techniques (e.g., Belkin, 2021; Mauzole, 2022) alongside the upwelling detection could be performed.

The clustering K-means model is currently fitted, and geophysical variables selected for, the SCU. This method can be expanded in future, either to the rest of the WIO or indeed additional seasonal upwelling systems globally, but with some adaptations. It is possible that some covariates other than Chl and SST become more important for successful clustering. This may be particularly relevant in, for example, regions where there is substantial riverine input, which is typically higher Chl at high SST, rather than the high Chl/low SST expected as an upwelling signature. Another example would be areas where primary production is not limited by nutrient availability (e.g., at higher latitudes); the addition of nutrients via upwelling in these areas may lead to a more muted response in Chl. Although potential extra remotely sensed covariates will still be limited by the available resolution, as in the case of SLA, model information may possibly be used to supplement remote sensing data.

An unsupervised ML clustering (K-means based) approach was used to detect and then classify seasonal upwelling surface signature off the Somali coast using satellite observations. This approach successfully delineates upwelling core and surrounds as well as non-upwelling ocean regions. The technique is shown to be temporally robust (seasonally and interannually) with accurate classification of NRT data not used in the fitting of the classification model. Once upwelling regions have been identified, classification of extreme upwelling events was performed using a threshold approach that includes confidence intervals derived from historical data. These approaches, that we call “Upwelling Watch”, are designed to be adaptable to meet users’ needs.

Due to the high productivity in seasonal coastal upwelling systems, such as analysed here in the WIO, they support a large volume of fishing activity sustaining food security in the region for millions of people (Taylor et al., 2019). At the same time, upwelling systems can be areas of reduced oxygen and more pronounce acidification (e.g., Kämpf and Chapman, 2016), negatively impacting marine ecosystems. Thus, extremes in upwelling along with general upwelling variability can directly affect fishing catch success rates and the health of marine ecosystems in general. This “Upwelling Watch” service, designed to detect these upwelling systems and their anomalous behaviour, can be very useful provided it is accompanied by schemes to raise awareness and maximise use rates for those most in need of it. Furthermore, local fisheries data collection and analysis needs to work hand-in-hand with this service to be able to identify and document biological consequences of the extreme upwelling events. An implementation scheme alongside the full development of the system could involve interactions with local management, governments, and NGOs in order to aid in communication of its benefits.

This work has shown promise within the SCU system with aims to expand it in future to the rest of the WIO using similar clustering techniques although recognising that different upwelling systems can provide their own unique problems, potentially requiring adaptations to both the method and geophysical information used. This “Upwelling Watch” method can aid fisheries management, by potential defining target areas for fishing. The “Upwelling Watch” method may also provide broader scientific value by providing a definition of upwelling regions, and classification of their modes of variability. This subset of data can then be further analysed in numerous ways. Examples might include studying temporal and spatial dynamics of upwelling, addressing potential questions such as whether upwelling is changing in its spatial extent as well as quantifying marine heatwaves and cold spells and associated extremes of Chl (green waves), in this and potentially other seasonal upwelling regions.

Publicly available datasets were analyzed in this study. This data can be found here: https://resources.marine.copernicus.eu/product-detail/OCEANCOLOUR_GLO_CHL_L4_NRT_OBSERVATIONS_009_033/INFORMATION https://resources.marine.copernicus.eu/product-detail/OCEANCOLOUR_GLO_CHL_L4_REP_OBSERVATIONS_009_082/INFORMATION https://resources.marine.copernicus.eu/product-detail/SST_GLO_SST_L4_NRT_OBSERVATIONS_010_001/INFORMATION https://resources.marine.copernicus.eu/product-detail/SEALEVEL_GLO_PHY_CLIMATE_L4_MY_008_057/INFORMATION https://resources.marine.copernicus.eu/product-detail/SEALEVEL_GLO_PHY_L4_NRT_OBSERVATIONS_008_046/INFORMATION

Methodology and investigations: MH. Writing—original draft preparation: MH. Writing—review and editing: all. Supervision: FJ, MS, EP. Project administration: EP. All authors contributed to the article and approved the submitted version.

This study was supported by the Global Challenges Research Fund (GCRF) under NERC grant NE/P021050/1 in the framework of the SOLSTICE-WIO project (https://www.solstice-wio.org/) as well as the UK National Capability project FOCUS (NE/X006271/1).

All data used in this study was sourced from the Copernicus Marine Service (CMEMS- marine.copernicus.eu). SST data was produced by the OSTIA project, Chl data by the Copernicus-GlobColour team, and SLA was processed by CMEMS. The authors would like to thank those involved for producing these datasets and for making them freely available.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.