We reviewed the existing literature on monitoring pelagic Sargassum through remote sensing techniques, using Boolean search in the Scopus database, for the period 1991 to date (March 23rd, 2023). We used the keywords “remote sensing” and “pelagic Sargassum” ((TITLE-ABS-KEY (remote AND sensing) AND TITLE-ABS-KEY (pelagic AND Sargassum). The search results provided 22 documents (20 research articles, 1 conference paper, and 1 note). Twenty of these documents were published between 2015 and 2023. We repeated the search, replacing the keywords “pelagic Sargassum” with “floating Sargassum”. In this case, we obtained 37 documents (35 research articles and 2 conference papers). Thirty-one of these documents were published between 2015 and 2023. The second search contains 29 documents that are different from those in the first search. A third search used the keywords “Sargassum” and “Deep Learning” yielded 16 documents (13 research articles, 1 conference paper, and 2 conference reviews). Eleven of these documents were new and had not been part of the first two searches. The outcome of these three searches is shown in Figure 1 . Peer-review publications related to monitoring Sargassum through remote sensing (RS) and deep learning (DL) techniques have increased significantly in the last five years.

The searches yielded a total of 62 different documents, which were analyzed using the VosViewer software (Van Eck and Waltman, 2010). The papers with more than 30 citations were chosen for display in Figure 2 , which shows the 16 documents, represented by circles and an abbreviated citation of each article for identification purposes. The diameters of the circles demonstrate the normalized number of citations of a document, which is equal to the number of citations the paper has divided by the average number of citations of all the documents published in the same year. The normalization accounts for the fact that older documents have had more time to receive citations than more recent documents. The distance between the two papers in the image indicates the closeness of the papers in terms of co-citation links. Generally, the closer two papers are to each other, the stronger their relatedness. The most robust co-citation links between papers are represented by lines (Van Eck and Waltman, 2010).

Without being exclusive and to provide a panorama of the benchmark studies related to the remote sensing of Sargassum, Table 1 lists the five papers with the highest values in the number of citations, number of links, and normalized number of citations.

The seminal studies of Gower et al. (2006); Hu (2009); Wang and Hu (2016) appear in the first two categories (number of citations and number of links). In contrast, more recent studies appear in the normalized number of citations criterion. The highest number of normalized citations corresponds to Arellano-Verdejo et al. (2019). This study applied DL methods to Sargassum monitoring for the first time.

Among the papers found in the Scopus database, we have discussed those that have been a significant reference in the observation, monitoring, or quantification of Sargassum and have served as an inspiration for the development of new methodologies. We also recognize that, within the limitations of this study, this selection of papers can be debatable, and some relevant studies may not have been considered. However, this study includes the most representative research findings regarding the algorithms used in pelagic Sargassum monitoring.

The Vegetation Indices (VI) are mathematical expressions that have an input of radiometric measurements of the vegetation used to evaluate the temporal and spatial variations of many vegetal biophysical parameters (Liu and Huete, 1995). These indices have been widely applied in RS applications using aerial and satellite platforms (Xue and Su, 2017). VI are effective algorithms for quantitative and qualitative evaluations of vegetation coverage, vigor, and growth dynamics, among other applications. VI were initially used for land-based observations and provided the basis for marine observations. This study addressed those used to observe massive Sargassum blooms.

Between 1972 and 1973, the United States of America developed an experiment to map regional vegetation conditions during the Great Plains growing season (Rouse et al., 1974). Vegetation conditions were measured in vast regions using Multispectral Scanner (MSS) data from the Earth Resources Technology Satellite 1 (ERTS-1, later renamed Landsat-1). The radiance values recorded in ERTS-1 spectral bands 5 (0.6-0.7μm) and 7 (0.8-1.1μm), corrected for the angle of the sun, were used to calculate the band ratio parameter (BRP), defined as the difference in the ERTS radiance value measured in Bands 5 and 7, divided by their sum, which shows a correlation with the green aboveground biomass in the grasslands (Rouse et al., 1974). The BRP parameter also contributed to the Normalized Difference Vegetation Index (NDVI), which is defined as (Eq. 1):

Where R_NIR and R_RED are the reflectance in the near-infrared (NIR) and red bands, respectively. Many applications have utilized NDVI for land coverage/land usage. Several scholars have reported applications of the NDVI concept in studying massive algal blooms (Prangsma and Roozekrans, 1989; Kahru et al., 1993; Hu and He, 2008). The NDVI can be used when there is a high contrast between Sargassum and the background in the images, and their values can be used as input to other indices to discriminate between Sargassum and emulsified oil, garbage, and different types of seaweeds (e.g., Trichodesmium, Syringodium, and Ulva) (Hu et al., 2015). On the other hand Hu (2009) used it as a reference method to compare against a Floating Algae Index (FAI) that he developed to estimate the abundance of Sargassum (Hu, 2009). However, it is sensitive to environmental variables and observing conditions (e.g., atmospheric influences, solar/viewing geometry, and sun glint) (Hu, 2009), which makes it difficult for quantitative analyses, because the absolute NDVI values can be changed, manual adjustments are necessary, limiting automation and large spatial-scale applications (Hu and He, 2008; Hu, 2009). Several indexes have been proposed to overcome these issues (Huete et al., 1999), including the Enhanced Vegetation Index (EVI) (Liu and Huete, 1995), which is a feedback-based index that corrects for the interactive canopy background and atmospheric influences by incorporating background adjustment and atmospheric resistance concepts. This enhanced soil and atmosphere-resistant vegetation index is defined in Eq. 2 (Huete et al., 1999):

Where R_NIR , R_RED , and R_BLUE are the reflectance in the NIR, RED, and BLUE bands, respectively. G is the gain factor to compensate for the aerosol effects, and C ₁, C ₂, and C ₃ are the independent coefficients of the pixels to compensate for the impact of the vegetation background.

Despite efforts, EVI remains sensitive to environmental variables and observation conditions. Therefore, developing more stable algorithms for these conditions is necessary. Also, given the dynamics of the marine environment, the monitoring of Sargassum in the sea and its accumulation on beaches requires periodic information to know the moment-to-moment status of the phenomenon, which is often unavailable, making the data provided by algorithms challenging to validate for marine coastal zones.

Reflectance band-ratio algorithms use spectral bands in the blue and green regions of the visible spectrum to estimate Chl_a concentrations in open oceans (Blondeau-Patissier et al., 2014). Examples include NASA’s ocean color algorithms, such as the the SeaWiFS OC4 (O’Reilly et al., 1998; O’Reilly et al., 2000) and MODIS OC3M (Campbell and Feng, 2005). Most reflectance band-ratio algorithms were designed for global applications in deep ocean waters (Odermatt et al., 2012). Therefore, it is inappropriate to generalize the parameterization of some algorithms to all ocean regions (e.g., Dierssen and Smith (2000); Sathyendranath et al. (2001); Claustre and Maritorena (2003); Volpe et al. (2007)). On the other hand, the use of blue-green spectral bands for specific detection of Chl-a in coastal waters is affected by the absorption signal of CDOM and Total Suspended Matter (TSM). Therefore, in coastal waters, the quality of Chl-a detection is significantly reduced and is often considered unreliable. A more detailed performance analysis of these algorithms can be found in Blondeau-Patissier et al. (2014).

On the other hand, if appropriate spectral bands are available, it is possible to use the Sargassum Index (SI) and the red/green band ratio to identify the presence of Sargassum among other species of algae and seagrasses, in which the color green prevails (Dierssen et al., 2015). Hence, the SI is defined as Eq.3:

Where “r3” is the wavelength of a local reflectance peak, and “r2” is the wavelength of a local reflectance trough. Using Portable Remote Imaging Spectrometer (PRISM) images at 1 m resolution, the Sargassum Index was used to effectively discriminate Sargassum from Syringodium seagrass (Dierssen et al., 2015).

The red/green band ratio is defined as Eq.4:

Where “green” has a wavelength near 555 nm. The reason to define this index is to differentiate Sargassum from Trichodesmium (Cyanobacteria) because they have opposite red/green ratios (Dierssen et al., 2015).

A simple NDVI index can effectively discriminate Sargassum from the background in an image with high contrast. However, line depth (LD) and SI are required to enhance NDVI output to differentiate Sargassum from Syringodium blooms, oil, or garbage. SI can also differentiate Sargassum from other floating materials in environments free of emulsified oil and debris (Dierssen et al., 2015). Table 2 shows the steps to discriminate Sargassum from other algae and seagrasses or floating materials using PRISM imagery.

A methodology that has inspired several studies to avoid oxygen absorption at 687 and 760 nm and water vapor absorption at 730 nm is known as the baseline algorithm (Abbott and Letelier, 1999). Gower and co-workers used aircraft-based sensors to test several channels as a baseline for calculating the Fluorescence Line Height algorithm (FLH). They varied the bandwidth and position and eventually developed a simple linear model using three bands (Borstad, 1985), the general form is presented in Eq. 5. The mathematical expression of this algorithm is also known as Spectral Band Difference Algorithm (SBDA).

Whereby R ₂ is the radiance or reflectance depending on the use of the input data, measured in the peak of λ ₂, R ₁ and R ₃ are the radiance or reflectance of the baseline wavelengths, λ ₁ and λ ₃. Finally, K is a constant (Palmer et al., 2015). The subscripts can be rewritten with the reflectance used according to the study (Eq. 6). Note that the subscripts are generic, and the associated wavelength depends on the index used. For details, see Table 3 .

To make the expression shorter, it is usually rewritten, as shown in Eq. 7 and 8.

Where:

All SBDA algorithms discussed in this study are defined as the difference between the reflectance or radiance (R) in the central waveband (λ ₂) - which corresponds to the maximum of the red-edge effect -and a linear baseline, drawn between surrounding bands (λ ₁ and λ ₃, respectively) (Gower et al., 2005; Hu, 2009; Wang and Hu, 2016). SBDA algorithms take advantage of the fact that absorption tends to vary more rapidly with wavelength than dispersion; two adjacent reflectance spectral bands may have similar backscatter properties but will differ significantly in absorption. Therefore, this absorption can be quantified by spectral difference, which these algorithms seek to take advantage of (Blondeau-Patissier et al., 2014).

The first observation of Sargassum from space in 2005 (Gower et al., 2006) was achieved using imagery from the Medium Resolution Imaging Spectrometer (MERIS), on the Envisat satellite launched by the European Space Agency (ESA), and from the Moderate Resolution Imaging Spectroradiometer (MODIS), launched on both the Terra and Aqua satellites by the National Aeronautics and Space Administration (NASA). MERIS imagery was used as input for the Maximum Chlorophyll Index (MCI), and MODIS imagery for the Fluorescence Line Height (FLH) MODIS product. FLH and MCI belong to SBDA algorithms and were the inspiration for the Floating Algae Index (FAI) (Hu, 2009) and the Alternative Floating Algae Index (AFAl) (Wang and Hu, 2016). The SBDA algorithms use band triplets from the visible (VIS), infra-red (IR), near infra-red (NIR) and shortwave infra-red (SWIR) spectral regions ( Table 3 ). The wavelengths are selected depending on the phenomenon monitored to set the index, so the algorithm is sensitive to Chl-a fluorescence, Chl-a concentration, or surface flowering (Blondeau-Patissier et al., 2014). The features of these algorithms are detailed below.

Bio-optical modeling is an essential tool for understanding the effects of dense algal concentrations on light absorption and backscatter coefficients (Blondeau-Patissier et al., 2014). To bio-optical models have meaningful applications, it is necessary to calibrate them with values from in situ measurements of the variables used in the model so that the coefficients used are the most adequate to represent the phenomena under study as accurately as possible. To our knowledge, only the study of Schamberger et al. (2022) sheds light on a bio-optical algorithm specifically designed to detect Sargassum. This study adapted the semi-analytical radiative transfer model (Lee et al., 1998), proposed by Descloitres et al. (2021), which included the submerged Sargassum in rafts. This model is called Sargassum Radiative Transfer (SRT). Like Lee’s et al. (1998) model, SRT can simulate surface reflectance using the bio-optical properties of the water column, bottom depth, and bottom composition as inputs (forward model). The model uses the following variables as inputs: Chl-a, the concentration of Non-Algal Particles (NAP), the Colored Dissolved Organic Matter (CDOM) absorption coefficient, the seabed reflectance, and the depth (z) (Schamberger et al., 2022). These parameters determine the Inherent Optical Properties (IOP) of the water, such as the absorption (a(λ)) and backscattering (bb(λ)) coefficients that are used to model the deep-water reflectance prsdp(λ).

In the context of Computer Science, ML is a branch of Artificial Intelligence (AI) (Minsky, 1961) that allows machines to learn without being expressly programmed for it. An ML algorithm can learn from the data (Goodfellow et al., 2016), which is an essential skill for making systems capable of identifying patterns to be able to make predictions. This technology is present in several applications (e.g., content recommendation on streaming platforms, speech recognition from virtual assistants, autonomous cars, search engines, medical diagnoses, or fraud detection). The contribution of ML is the paradigm shift from rule-based programming to autonomous learning from data (Goodfellow et al., 2016). Additionally, ML methods are widely used to extract patterns and insights from the ever-increasing data streams in sensory systems (Camps-Valls et al., 2021). In general terms, ML helps to make a prediction or data grouping based on input data. Prediction problems can be divided into two broad categories: Regression problems, where the variable to be predicted is numerical, and Classification problems, where the variable to be predicted is part of a predefined set of categories. For the training of an algorithm, ideally, a balanced dataset is required, i.e., with homogeneous values of the different elements or variables included in the study case.

A characteristic of ML is that there are metrics for evaluating models and their performance quantitatively, allowing us to know the confidence level of the algorithm in different scenarios. A fraction of the dataset (generally 80%) is used as the training dataset, and the remainder is the test dataset, used to evaluate the algorithm’s performance. The ML evaluation metrics depend on the model type (regression or classification). Classification models are applied in Sargassum monitoring, and their most straightforward evaluation metric is accuracy, which is the percentage of observations correctly predicted by the model. However, it must be interpreted cautiously, especially if the classes to be predicted are out of balance (Hossin and Sulaiman, 2015).

The confusion matrix is another tool to evaluate and compare models. It compares the actual values of the target variable against the predicted values. The matrix number of rows and columns depends on the number of possible outcomes. There are only two possible outcomes for binary classification, so only two columns and two rows exist. The labels that make up a confusion matrix are true positive (TP), false negative (FN), false positive (FP), and true negative (TN). Additionally, the information from the confusion matrix makes it possible to calculate the following metrics: Precision Eq.9 and Recall Eq.10, as defined below (Hossin and Sulaiman, 2015).

Precision and Recall are used to define the metric F1 score as shown in Eq.11

The metrics above allow us to quantitatively determine the performance of the algorithm used for a particular study and compare different algorithms with each other (Hossin and Sulaiman, 2015). Then, we will describe innovative study cases that use ML techniques to improve the Sargassum monitoring in the ocean, the coast, and the beach.

The democratization of DL has been possible thanks to the availability of large datasets, the advances in hardware and parallelization, computational resources, and community efforts to develop and share codes. This has resulted in an increasing number of people using DL to solve different challenges. Nowadays, DL seems to be consolidating as a widely used scientific research paradigm (Camps-Valls et al., 2021). The vast amount of data about the Earth and its environments, generated by satellite observation platforms and other remote sensors has enabled Earth observation to join the data revolution. The use of DL in Earth Sciences has grown exponentially in recent years because the datasets available offer enormous potential to study terrestrial ecosystems and to address major societal challenges related to emerging issues (e.g., food, water, energy security, and climate change) (Camps-Valls et al., 2021). DL allows the analysis of RS data from another perspective. Below we describe some of the most recent studies incorporating DL to improve Sargassum observation.

In 2021, Rutten et al. (2021) characterized the temporal variation in Sargassum found in the reef lagoon in Puerto Morelos, Quintana Roo, Mexico. The study analyzes images taken every hour for approximately 5.2 years (September 2015 to November 2022) to observe the relationship between landings with wave energy, water level, and wind speed. With this information, the authors proposed a model to explain how the natural cleaning process of the Sargassum accumulated on the beach occurs. This study used a SVM to classify the images automatically. However, it is necessary to know more details of the algorithm adjustment process to reproduce the experiment; for example, the kernel used, the implication of training the algorithm with an unbalanced dataset, and confusion matrices for different cases. This study contributes to the knowledge about Sargassum dynamics and the capacity of the coastal system to remove Sargassum naturally on beaches.

Sargassum monitoring on beaches using photographs and videos has also been carried out through various citizen science initiatives, including Crowdsourcing. The increasing use of mobile devices, such as smartphones and tablets, and the infrastructure that interconnects them with the Internet, have made collective participation a vital source for photographs of the presence of Sargassum on beaches. Citizen science has recently been used as a methodology to collect field information for the validation and calibration of supervised algorithms for identifying Sargassum on the beaches (Arellano-Verdejo and Lazcano-Hernandez, 2020; Iporac et al., 2020). In addition, the study by Putman et al. (2023) suggests the inclusion of wind speed and citizen science data to improve Sargassum inundation reports along beaches. This study constructed its dataset by vertically cropping each photograph to remove the sky and then superimposing 50 random points on the image using the Coral Point Count with Excel extensions (CPCe) software (Kohler and Gill, 2006). Each point was classified into six classes (Sargassum, sand, sky, water, vegetation, or ‘other’). To estimate the relative amount of Sargassum present in the photo, a ratio was calculated between the number of Sargassum pixels and the sum of sand, water, vegetation, and Sargassum points.

The study conducted by Arellano-Verdejo and Lazcano-Hernandez (2021) proposed a methodology for monitoring Sargassum on beaches based on Crowdsourcing for capturing geotagged photographs. An application for mobile devices (e.g., smartphones, tablets) was developed so that volunteers could capture and send the photographs to the cloud for storage. A dataset of 2400 photographs (1200 with Sargassum and 1200 without) was built to train, test, and validate the classification algorithms. DL algorithms classified the images automatically into two classes (with and without Sargassum), and maps were created with GIS to visualize the presence or absence of Sargassum on beaches. To accomplish this task, three state-of-the-art CNNs (LeNet-5, AlexNet, and VGG16) were chosen for image classification and compared on their performance to find the most suitable architecture. Each CNNS was adjusted for the classification process to use the new dataset. LeNet-5 was modified to work with RGB images. The last layer of 1,000 categories in the AlexNet and the VGG16 architectures was replaced by a fully-connected layer for two classes (Sargassum and non-Sargassum images). AlexNet and VGG used augmented data and transferred learning to maximize the generalization capacity of the network for the given dataset. The architecture with the best performance in this study was AlexNet, with an f1 score of 92%. This is a significant result considering that the training was performed with a relatively small set of images (2400). The main contribution of this study is the building of the first dataset of segmented and geotagged images that can be used in other studies. This is important since there is a lack of datasets for studying natural phenomena. Additionally, this study uses the results of the proposed methodology to develop presence/absence maps of Sargassum along the beaches.

The study conducted by Santos-Romero et al. (2022) proposed a method to classify images automatically to build adequate datasets for studies that highlight the presence of Sargassum along the beaches. The images were classified into two classes, in-depth linear perspective and no linear perspective. The Transfer Learning technique on three Artificial Neural Networks (ANN’s) from the literature (ResNet50, MobileNetv2, and VGG16) was implemented to find the optimal architecture. The study also used the Collective View dataset as input for constructing its training dataset. From that collection, 5,000 pictures were manually selected and classified into two balance classes: 2,500 images with a linear beach perspective and 2,500 images without this perspective. The dataset was divided into two subsets using an 80-20 hold-out scheme (Chollet, 2021). 80% of the data were used for training the model and 20% for its testing. The ANN’s performance was assessed with the Accuracy, Precision, and F1 score metrics. A resampling analysis of the F1-Score metric was performed with the Bootstrapping technique, concluding that VGG16 and MobileNetV2 obtained similar results with a confidence interval of 95% between 0.88 and 0.92. Finally, considering both architectures’ spatial and temporal complexity, MobileNetV2 was proposed as the most appropriate for classifying images with and without linear perspective in depth. One of the main contributions of this study is the implementation of resampling to analyze the algorithm’s performance. Resampling is a statistical technique that allows computing estimates of the variance, bias, and confidence intervals of an estimator, using random subsets of the dataset.

The study by Valentini and Balouin (2020) used a SOLARCAM system for beach monitoring, to compute the semantic segmentation of the imagery, the following was carried out: 1) the image was partitioned in superpixels using the sticky-edge adhesive algorithm; 2) the pre-trained CNN MobileNet-V2 classified the superpixels; and 3) semantic image segmentation was performed through fully connected Conditional Random Field (CRF). The segmented classes were algae, anthropic structure, seafoam and swash, sand, sky, vegetation, and seawater. According to the F1 score metric, the performance of the proposed methodology related to the presence of Sargassum reached a maximum of 90%. At publication, this study was the first to apply the semantic segmentation technique to beach monitoring. The proposal is scalable and would only need to integrate the results into a map for visualization.

Balado et al. (2021) applied a semantic segmentation technique to classify imagery of five macroalgal species (Bifurcaria bifurcata Linnaeus, Cystoseira tamariscifolia (Hudson) Papenfuss, Sargassum muticum (Yendo) Fensholt, Sacchoriza polyschides (Lightfoot) Batters, and Codium spp). The study used data augmentation to build the high-resolution dataset to retrain three CNNs and compare their performance (Mobilnet-V2, Resnet 18, and Xception). The framework used in this study was DeepLabV3 in Matlab. In total, 130 images were labeled and distributed: 90 images for training, 10 for validation, and 30 for testing. This paper shares the hyperparameters used in the learning process of the algorithm (optimization method, learning rate, momentum, L2 regularization, max epochs, and batch size), which are useful for conducting a replica of the study and is information that has usually been omitted in papers. This is one of the pioneering applications of semantic segmentation to differentiate five species of macroalgae.

To improve the results of this type of study, many ground images are required, as well as a good segmentation of the training dataset. On the other hand, geospatial metadata is required to apply the algorithm in the elaboration of maps or other useable products in decision-making. To implement this proposal in aerial images, new datasets are needed to train, test, and validate the algorithm. Additionally, the number of classes required for segmentation needs to be analyzed since UAV-acquired imagery includes information that does not appear in ground images.

In 2022, Arellano-Verdejo et al. (2022), used the collection of geotagged photographs of Arellano-Verdejo and Lazcano-Hernandez (2021) to apply a semantic segmentation technique to propose a new mapping methodology for estimating Sargassum coverage on beaches. Implementing the semantic segmentation image technique requires several steps, one of which is to build a dataset consisting of pairs of pictures (source and segmented). To achieve image segmentation, two stages are required: image translation and subsequent segmentation. Image translation was carried out using the Pix2Pix architecture, and segmentation was performed using the k-means algorithm. All images in the dataset were segmented into three classes (Sargassum, sand, and others). If all the Sargassum observed in the photographs lies on the beach, the total beach area was confirmed by Sargassum and beach pixels. With this criterion, the Sargassum coverage percentage over the beach on every picture was calculated for each segmented photograph. The number of photographs available for the study zone allowed statistically reliable mapping of the percentage of Sargassum coverage from the segmented images. As an element of validation of the proposed methodology, an orthophoto generated by a UAV flight of the study area was used for one of the photographic sampling dates. The main contributions of this study were the generation of the first geotagged and segmented image dataset for Sargassum monitoring for the study area (Mahahual, Quintana Roo, Mexico), and the design of a methodology to generate maps of Sargassum coverage on beaches, which provides more information than maps of the presence/absence of Sargassum. One challenge of the methodology is the need for a constant flow of images for map creation. This situation is generally not possible since society’s participation rate is overall low.

The evolution of Sargassum observation and study approaches can be summarized chronologically, methodologically, by area, and by scale.

Different platforms provide information on Sargassum biomass in the ocean, coastal waters, and beaches. However, ideally, they should provide consolidated information for all environments. The only source of periodic reports on the distribution and abundance of Sargassum in the Atlantic Ocean is the Satellite-based Sargassum Watch System (SaWS) from the Optical Oceanography Laboratory of the University of South Florida ² . Most countries affected by Sargassum in the Greater Caribbean rely on this information to estimate landing volumes and develop management strategies. Nevertheless, adequate measurements of Sargassum coverage and biomass in coastal waters and on beaches are needed to understand spatial and temporal variability in landings, both within and between countries, and to feed numerical prediction models. Thus, to improve Sargassum’s understanding and management, the current geoportals and information sources must evolve into hybrid Sargassum monitoring observation and survey systems.

Sargassum monitoring in coastal waters faces several challenges. In the CS, the shallow oligotrophic waters and enormous ecological diversity result in a wide variety of colors, making it difficult to correctly identify the elements in a scene. Thus, monitoring coastal areas requires images with a sharp spatial resolution (1 m) and high revisit periods, preferably once a day. Although satellite platforms that offer very high spatial resolution images do exist ³ , they are economically inaccessible for most researchers in the region.

A significant research gap has been quantifying Sargassum landing volumes because in situ measurements are difficult and time-consuming (Rodríguez-Martínez et al., 2022). The computer vision algorithms for the classification and semantic segmentation of Sargassum in photographs collected by citizen science are a step towards developing beach monitoring scalable techniques. Nevertheless, methods to discriminate Sargassum from other types of macroalgae or seagrasses are still needed to reduce false positives and avoid overestimating Sargassum landing volumes. The main challenge of monitoring techniques based on citizen science is for society to adopt them. The platforms based on data and images captured by the voluntary participation of citizens depend on a constant flow of data to produce value-added products. Also, for these platforms to be successful and use their total capacity, they should be run and promoted by the institutions in charge of Sargassum management. To date, most of the information available consists of photographs and isolated data records in different formats and with little or no knowledge of metadata.

Incorporating ML methodologies in the Sargassum observation process has contributed to minimizing the background noise in the satellite images, thereby reducing the number of false positives and negatives (Wang and Hu, 2020; Wang and Hu, 2021). However, it requires being more rigorous in its implementation, and evaluation. There are well-defined steps in ML that must be followed to successfully implement an algorithm. However, they have either not been followed or have not been discussed in the studies. For example, the supervised classification algorithms applied for Sargassum monitoring, requires balanced datasets with thousands of images for training stage (training, testing, and validation). However, most of the studies that employ ML algorithms, lack a detailed explanation about the characteristics of the dataset and about the hyperparameters settings. There is also a need to implement confusion matrices to analyses algorithm performance and implement metrics (F1 score, IoU), to assess with more detail the algorithm performance. Table 4 shows the general aspects expected to be described in any study using supervised classification algorithms.

Regarding the application of high-resolution satellite imagery for monitoring Sargassum on the coast and beaches also has potential. Although its implementation may be economically costly, its application may be feasible for observing strategic study areas. On the other hand, the development of methodologies using inputs from geostationary satellite platforms is also pending.

Regarding citizen participation in monitoring proposals, an action that we consider essential for gathering data is the coordination and periodic participation of several sectors of the citizenry to serve their purpose.

Regarding the application of ML in the monitoring of Sargassum we consider is still in the initial stage, their potential has not been used yet, and improvements will help to obtain more information from satellite imagery. The current challenges are adequately implementing image processing and computer vision techniques to eliminate noise in the images regardless of their scale and to later ingest the image to the preferred index to detect Sargassum with the least possible error. We believe that designs from scratch are needed for Sargassum discrimination. On the other hand, DL are not always necessary, image processing and computer vision techniques in some cases may be sufficient for the Sargassum discrimination in an image.

A common feature in studies that use ML as a methodology to enhance Sargassum contrast in images is the use of the UNet architecture, which is a convolutional neural network developed for biomedical image segmentation at the Department of Informatics of the University of Freiburg (Ronneberger et al., 2015). One of the main features of this architecture is that it offers useful image translations from small training data sets (i.e., hundreds of images). This feature has allowed its application to address problems where data is scarce. Such is the case of Sargassum observation and monitoring, where the UNet has been used as an architecture base to improve Sargassum observations in oceans, coasts, and beaches.

Another challenge is the design of hybrid systems that incorporate information at different scales and environments to improve our understanding of the Sargassum phenomenon and the development of better strategies for its management, valorization, and disposal.