Locally developed models improve the accuracy of remotely assessed metrics as a rapid tool to classify sandy beach morphodynamics

Study area

This study was carried out on beaches located on the northern coast of the State of São Paulo, Brazil. This area was chosen due to the heterogeneity of shore morphological features, with a more straight, exposed shore on the southern part, and a sinuous, more sheltered coastline towards the northern part of the study area (Fig. 1). This characteristic was a desirable feature, as it allowed us to evaluate the efficacy of the method on a large range of morphodynamic types, from dissipative to reflective beaches, also including exposed and sheltered shores. Additionally, this area involves a complex environmental and governance setting, as much of the coastal region is located within protected areas, but suffers from threats from urbanization and intense tourism (Pierri Daunt et al., 2021). With approximately 200 beaches located along ∼200 km of coastline, many of these being remote and having a difficult access, methods that can provide a rapid assessment of physical and biological characteristics of beaches, as well as providing baselines for their monitoring, are valuable tools for the management of these ecosystems.

A total of 29 beaches were sampled to identify the environmental characteristics (i.e., in-situ morphodynamic and hydrodynamic characteristics) between February and June/2019. We divided each beach into three subareas (the middle region and both beach corners) with a minimum distance of 100 m among them, which were further treated as individual sites (n = 87). This division was done due to the great alongshore variation in natural features (i.e., waves, beach slope) which affect beach type. Additionally, human activities in beaches may be directed towards specific areas, especially in extensive beaches (Alexandrakis, Manasakis & Kampanis, 2015; Machado et al., 2016), and management may benefit from considering the local variability within beaches.

Assessment of beach physical characteristics

To evaluate beach morphodynamics in situ, beach profile and slope (inclination of the intertidal), and granulometric parameters were assessed at each sampling site. Beach profile was measured from the supratidal to the end of the swash zone following the method described by Emery (1961), during low spring tides. The width of the intertidal area was registered in the profiling and the slope, calculated using the trigonometric relationships, estimated as 1/tg(β). To estimate the mean grain size, five sediment samples were collected along the beach profile. In the laboratory, after drying, they were sieved into twelve granulometric fractions and the mean grain calculated based on Folk & Ward (1957). Sites were classified into five morphodynamic types: dissipative, intermediate-dissipative, intermediate, intermediate-reflective, and reflective. This characterization was based on the environmental features measured at each site and further supported by the calculation of the Beach Index (BI, McLachlan & Dorvlo, 2005).

Remote assessment of beach characteristics

Remote assessment of beach conditions was done by measuring specific features on satellite images. Images were obtained from the Google Earth Software (CNES/Airbus, Maxar Technologies), as it is the most user-friendly and accessible source, whose application can be standardized across regions, facilitating a rapid and easily applicable assessment by a wider range of users. The date of the available images was chosen as close as possible to the date of sampling of each beach.

Similar to Harris, Nel & Schoeman (2011), the following characteristics were assessed: beach intertidal width, surf-zone width, number of waves and bores in the surf-zone, and degree of exposure (Table 1, Fig. 2). All these metrics are expected to change within the beach morphodynamic continuum and are thus considered good proxies for classifying beach type. Beach intertidal width was measured as the distance from the drift line to the lowest position of the swash. Surf-zone width was measured as the distance from the innermost to the outermost swash edge. The number of waves and bores within the surf-zone was determined by distinguishing the number of visible ripples. Surf-zone type, a variable included in Harris, Nel & Schoeman (2011), was initially considered, but we found this variable to be difficult to identify in many instances, especially in standardizing across a set of temporal images, and thus we excluded this variable from the analysis. Finally, the degree of exposure was coded from 0 (very sheltered) to 4 (very exposed) (See Fig. S1 for image distinctions among types). The degree of exposure was established by observing the shape and degree of curvature of the headlands, as well as the presence of nearshore islands or other formations that could affect wave exposure (Turnbull et al., 2018).

We used two methods to assess the value of each characteristic. As performed by Harris, Nel & Schoeman (2011), we first measured environmental characteristics from a single image taken on the closest day available to the date of the in-situ beach sampling. However, beach characteristics, such as tide and surf-zone dynamics, are susceptible to daily variability brought by climatic and oceanography variables, such as wind direction and intensity, and storms-driven changes in the surf-zone climate (Ortega et al., 2013; Roberts, Wang & Puleo, 2013). For this reason, we also estimated the value of beach characteristics based on the mean value obtained from five satellite images, using the images closest to the day of the in-situ sampling. This dual approach of estimating beach characteristics allows us to compare if temporal variations can affect the accuracy of the method and also evaluate which metric is most susceptible to this temporal variability. We avoided images that were visibly taken during high tide, as an estimation of intertidal parameters would be misleading in such instances, and used images where the intertidal area was clearly exposed. Additionally, images where low quality or conditions (e.g., very pixelated, taken during cloudy conditions) hindered a proper measurement of beach characteristics were not included in the assessment.

This procedure was carried out for all 87 sites sampled in the 29 target beaches, and this data was used in further analyses to build the predictive model (train dataset, Fig. 3). To test the predictive model, we also processed satellite images for 30 additional beaches from the study area (n = 77 beach sites, as some beaches were small and were considered as a single unit—beaches with less than 100 m long—or divided into only two sites –beaches with less than 300 m long) (test dataset, Fig. 3). The morphodynamic type of those beaches was obtained from literature reports (e.g., Amaral & Denadai, 2011; Souza, 2012; Checon et al., 2018), with the support from the slope and sediment grain size data available from a partner project (see resulting classification on Table S2). This train dataset was used to test the robustness of the model developed with the training dataset in predicting beach type. Finally, to extrapolate the results from the predictive model to the whole North Coast of São Paulo and derive an overall characterization of beach morphodynamics, we carried out the satellite image processing for the remaining continental beaches of the study area, which were divided into sites according to the beach length (n = 73 beaches, 144 sites). We only excluded severely squeezed beaches, which are present in some areas, as the intertidal cannot be properly estimated. This third dataset (extrapolation dataset, Fig. 3) was used only to predict the beach types based on the results from the model, provided the model was robust enough to support this classification.

Data analyses

Environmental and remote characterization

The satellite-based measurements of intertidal width were strongly correlated to the in-situ measurements, for both point (r = 0.876) and mean values (r = 0.931), indicating the suitability of this approach as a proxy of variables related with width. The correlation, however, was stronger for mean-based values (CI ≠ 0, range = −0.098; −0.031), which were therefore applied in further analyses. Conversely, satellite-based intertidal heights were very poor proxies of field measurements, as they generally overestimated the height of wide and flat profiled beaches, and underestimated the slope of short, steep profiled beaches. For this reason, correlation coefficients were low and negative (point: r = −0.204, mean: r = −0.159), regardless of the method (CI = 0, range = −0.124; 0.033). Due to this very low correlation with in-situ values, we dropped the intertidal height from further analyses.

The environmental characterization showed that sampled sites comprehended a large portion of the recognized morphodynamic gradient of beaches, with beaches ranging from fine to coarse sands, from gentle to steep slopes (Table 1). From the 87 beach sites, 18 were classified as dissipative, 17 as dissipative-intermediate, 22 as intermediate, 11 as reflective-intermediate and 19 as reflective beaches (Table 1). Although for most beaches the same morphological beach type was classified among its sites, a third of the beaches had variations, going from intermediate to dissipative (e.g., Lagoinha, Cidade) and from intermediate to reflective (e.g., Una, Sahy) (Fig. S2). The principal component analysis showed that the sites followed a clear distribution along the morphodynamic gradient evidenced by the first axis (PC 1 = 84.07%), characterized by a gradient of variation in the proxies of beach type (e.g., mean sediment grain size, beach slope, intertidal width) (Fig. 4).

Model building and validation

The conditional inference tree model built with the 87 sites dataset showed that intertidal width and degree of exposure were the only variables selected to predict beach type (Fig. 5). Four terminal nodes were identified: Node 7 was characterized by beaches with intertidal areas >57.63 m, which classified dissipative beaches with a small error (5.0%). Node 6 was characterized by beaches with intertidal areas between 57.63 and 34.85 m, which classified most intermediate dissipative beaches (error = 12.5%). Node 4 was characterized by beaches with intertidal areas <34.85 m and having exposure ≤ 2 (i.e., moderately exposed, sheltered and very sheltered conditions) and enclosed most of the intermediate beaches (error = 10%). Node 5 was characterized by beaches with intertidal areas <34.85 m, but having exposure >2 (i.e., exposed and very exposed beaches). This node was composed mostly of reflective beaches; however, there was a higher error rate (38.8%), since all intermediate reflective beaches were also classified within this node. No reflective or intermediate reflective beaches were classified outside this node. For this reason, we grouped the two beach types within this node, which reduced the error to 7.5% (Fig. 5). The final classification was as follows:

• Dissipative beaches: Intertidal Width >57.63 m (error = 5.0%);

• Intermediate Dissipative beaches: 34.85 m <Intertidal Width <57.63 m (error = 12.5%);

• Intermediate beaches: Intertidal Width <34.85 m; Exposure ≤ 2 (error = 10.0%);

• Intermediate Reflective/Reflective beaches: Intertidal Width <34.85, Exposure >2 (error = 7.5%).

The estimation of the accuracy of this classification through internal validation provided good results. When the classification and prediction datasets were taken as subsets from the initial dataset, the percentage of correct classification (in comparison with the classification based on field measurements) was 92.8%, the weighted kappa indicated a very good agreement (k = 0.897) (Table 2). The confusion matrix shows that only two sites were misclassified using this method (i.e., intermediate beaches classified as intermediate-dissipative) (Table 3). Mean sensitivity and specificity were both high, 92.8 and 97.5% respectively (Table 3). Sensitivity was 100% for all categories, except the intermediate type (70%). Specificity was 100% for all categories, except for intermediate-dissipative (90%).

When testing the accuracy of the classification scheme to predict the beach type of the 77 sites of the test dataset for external validation, the percentage of correct classification was even higher, reaching 94.8% of agreement. The weighted kappa indicated a very good agreement between predicted and observed morphodynamic types (k = 0.907) (Table 2). In the confusion matrix, five of the 77 sites were misclassified and these mismatches were evenly distributed among the categories, with most occurring among close morphodynamic types, except for a single intermediate-reflective/reflective site which was predicted to be intermediate dissipative (Table 3). Mean sensitivity and specificity were also high, 95 and 98.1% respectively (Table 2). Sensitivity was 100% for all categories, except the intermediate-dissipative (80%). Specificity was 100% for all categories, except for intermediate-reflective/reflective (98%).

Finally, when predicting the beach type based on the previous classification scheme developed for the South African coast, the accuracy was very low. Only a third of the sites were correctly classified (PCC = 34.8%), and the weighted kappa indicates a poor agreement (k = 0.019) (Table 2). The confusion matrix shows that 53 of the 77 sites were misclassified by the predictive model (Table 3). Sensitivity was higher for predicting the intermediate-reflective/reflective type (67.7%) but was low for dissipative (33.3%) and intermediate-dissipative types (35.0%), and very low for the intermediate type (3.3%), where only a single site was correctly classified. For the latter category, most sites were erroneously classified as intermediate-reflective/reflective type (Table 3). Specific values were higher than sensitivity, being 100% for dissipative and 92.9% for intermediate-dissipative but having lower values for intermediate (57.4%) and intermediate-reflective/reflective sites (48.2%). Mean sensitivity and specificity values were 34.8% and 74.6%, respectively (Table 2).

Model extrapolation

The extrapolation to the whole study area allowed us to provide a characterization of the 310 beach sites along 147 beaches of the study area (Fig. 6). Considering the 4 beach categories delimited by the model, most sites were classified as intermediate (n = 134, 43.5%). Although the second most common type was intermediate reflective/reflective (n = 76, 24.6%), which includes two beach types due to the inability of the model to differentiate among them. The intermediate dissipative and dissipative types were assigned to 53 (17.2%) and 45 (14.6%) sites, respectively. There is a clear distinction in beach type distribution along the study area. At the southernmost portion the coast is more exposed, and although all beach types are found, this region is where intermediate reflective/reflective beaches sites are most predominant, while more dissipative types are found towards the São Sebastião Channel (Fig. 6A). At the northern area of the São Sebastião municipality, beaches are heavily sheltered by the São Sebastião Island, and most sites are classified as intermediate (Fig. 6B). The central part of the study area is also divided in two; the southern is heavily sheltered inside the Caraguatatuba Bay, and most beaches are of intermediate dissipative and dissipative types, and the northern area is a more exposed shore, with most beaches belonging to intermediate and intermediate reflective/reflective type (Fig. 6B). The coast of the municipality of Ubatuba is a very heterogeneous coast, with the highest number of beaches, due to the heavily jagged coast, especially in the southern area (Fig. 6C). This result in a mosaic of beach sites of distinct types very close to each other, with sheltered beach sites within embayments, ranging from intermediate to dissipative type, and exposed beach sites on the outside of these embayments, with beaches ranging from intermediate to intermediate reflective/reflective type. At the northernmost portion of the study area, the coast is less jagged, and beaches are more exposed, with most sites classified either as intermediate reflective/reflective and dissipative types (Fig. 6C).

Implications for management

The method applied here was very accurate in detecting beach morphodynamic type, which can be a valuable tool to characterize coastal areas with many beaches with remote or difficult access, as well as reduce costs with field samplings to assess and monitor relevant beach characteristics. Furthermore, identifying the morphodynamic type of beaches is an important aspect for management purposes, as this metric is highly related to biodiversity patterns (Defeo & McLachlan, 2005), to the response and recovery from impacts, and also have direct repercussions on the quality and quantity of ecosystem services (McLachlan et al., 2013; Harris et al., 2014; McLachlan & Defeo, 2017).

For instance, the impact of organic and inorganic pollution, a recurrent problem in coastal areas worldwide, is known to be stronger in dissipative beaches, especially in sheltered conditions, due to the slower recovery linked to the lower permeability of the beach face (De La Huz et al., 2005; Harris et al., 2015). In this regard, the knowledge of beach type coupled with information of potential impacts, such as the risk of oil spills and/or presence of sewage outfalls in the beach, common problems in our study area (Zanardi-Lamardo, Bícego & Weber, 2013; Santos & Turra, 2017), can significantly help managers to predict and minimize impacts on beaches and their biodiversity. Plastic pollution, one of the most pressing problems in the marine environment, representing up to 80% of marine waste (Auta, Emenike & Fauziah, 2017), is also suggested by recent studies to have its accumulation linked to the energy and morophodynamic of beaches (Tsukada et al., 2021; Wilson et al., 2021). Thus, there is great potential for integrating morphodynamic types and the dynamics of litter on beaches to support management strategies (Fanini et al., 2021).

Aside from classifying beach morphodynamics, the measurements of beach features can be a rapid way of assessing information relevant for other management and economic purposes. For instance, the intertidal width, a metric with a strong correlation between local and satellite-based measurements, can be used to infer the suitability of beaches for tourism activities. Studies assessing tourist preferences found that beaches with large intertidal areas are usually preferred by general tourists, likely due to the available space, as well as the smooth slope commonly associated with wide beaches, which facilitates the practice of recreational activities (Onofri & Nunes, 2013). In contrast, surfers prefer beaches with a higher number of waves and wider surf-zones (Philips & House, 2009), metrics that were easily assessed in the present study.

Additionally, beaches with restricted supralittoral areas (i.e., with the intertidal located too close to the upper beach limits), especially those with armoring structures, may be more susceptible to climate change impacts, due to being more prone to rises in the mean sea level (Pontee, 2013; Leo et al., 2019). Thus, if beach width metrics are coupled with information of anthropic structures in beaches, then susceptibility to climate change impacts could be inferred. Remote imaging has also the potential to estimate features characteristic of urbanization such as shoreline occupation, suppression of vegetation, and presence of coastal armoring structures (Hall & Hossain, 2020; Morgan et al., 2022), which could be incorporated into an integrated framework with morphodynamic classification to assess this vulnerability.

Finally, the corroboration of the correlation between remotely measured metrics with in-situ characteristics represents an opportunity to use this tool to support predictive models of biodiversity, such as habitat suitability modeling. The use of remote sensing to aid niche modeling to predict and map biodiversity is common in terrestrial ecological studies (Goetz et al., 2010; Leitão & Santos, 2019; Wang & Gamon, 2019). However, its use to predict biodiversity in beach ecosystems is very limited (e.g., Marzialetti et al., 2021), and most remote sensing tools in coastal studies are used to assess and monitor physical characteristics (Kroon et al., 2007; Mars & Houseknecht, 2007; Luijendijk et al., 2018). Beach morphodynamic characteristics are considered important predictors of biodiversity, both from benthic assemblages, as well as surf-zone fishes (Defeo & McLachlan, 2005; McLachlan & Dorvlo, 2005; Shah Esmaeili et al., 2021). Evaluation of the degree of exposure may also be used to predict ecological processes, such as nursery potential (Oliveira & Pessanha, 2014) and biological invasions (Hampton & Griffiths, 2017). For this reason, further studies could aim to uncover the potential of the remote sensing approach in contributing to modeling biodiversity and sustainable use in sandy beaches.