Next Article in Journal
Organic Geochemical Signatures of the Upper Miocene (Tortonian—Messinian) Sedimentary Succession Onshore Crete Island, Greece: Implications for Hydrocarbon Prospectivity
Next Article in Special Issue
Assessing the Reliability of a New One-Line Model for Predicting Shoreline Evolution with Impoundment Field Experiment Data
Previous Article in Journal
Impact of Design Parameters on the Dynamic Response and Fatigue of Offshore Jacket Foundations
Previous Article in Special Issue
Critique on Ecological Methodologies Used in Water Quality Studies and Coastal Management: A Review
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:

An Integrated Monitoring System for Coastal and Riparian Areas Based on Remote Sensing and Machine Learning

Centre for Research and Technology Hellas, Information Technologies Institute, 57001 Thessaloniki, Greece
Geosystems Hellas S.A., 11632 Athens, Greece
Laboratory of Environmental Quality and Geospatial Applications, Department of Marine Sciences, University of the Aegean, 81100 Mytilene, Greece
Author to whom correspondence should be addressed.
J. Mar. Sci. Eng. 2022, 10(9), 1322;
Received: 30 August 2022 / Revised: 8 September 2022 / Accepted: 13 September 2022 / Published: 18 September 2022
(This article belongs to the Special Issue Decision Support Systems and Tools in Coastal Areas)


Nowadays, coastal areas are exposed to multiple hazards of increasing severity, such as coastal floods, erosion, subsidence due to a combination of natural and anthropogenic factors, including climate change and urbanisation. In order to cope with these challenges, new remote sensing monitoring solutions are required that are based on knowledge extraction and state of the art machine learning solutions that provide insights into the related physical mechanisms and allow the creation of innovative Decision Support Tools for managing authorities. In this paper, a novel user-friendly monitoring system is presented, based on state-of-the-art remote sensing and machine learning approaches. It uses processes for collecting and analysing data from various heterogeneous sources (satellite, in-situ, and other auxiliary data) for monitoring land cover and land use changes, coastline changes soil erosion, land deformations, and sea/ground water level. A rule-based Decision Support System (DSS) will be developed to evaluate changes over time and create alerts when needed. Finally, a WebGIS interface allows end-users to access and visualize information from the system. Experimental results deriving from various datasets are provided to assess the performance of the proposed system, which is implemented within the EPIPELAGIC bilateral Greece-China project. The system is currently being installed in the Greek case study area, namely Thermaikos Gulf in Thessaloniki, Greece.

1. Introduction

Coastal and riparian zones are the link between land and sea and are unique areas with excellent biodiversity and ecosystems [1]; they are valuable from an environmental, social, and economic point of view. The European Commission, as well as many other political and environmental organizations, have issued a series of guidelines and practices regarding the protection of these areas with emphasis on climate change, which will greatly affect these zones with the risk of even their extinction [2].
Modern approaches in Earth Observation (EO) and Remote Sensing fuse multi-temporal data from different sensors for monitoring these areas. Specifically, in order to design and implement useful and easy to use monitoring systems, information regarding different problems and risks that may affect coastal areas needs to be collected and analysed. Remote sensing and space technologies have valuable resources for obtaining such information, while analysis and modelling, including robust machine learning techniques, can be used to extract additional knowledge and key insights. An extensive review of successful application for coastal management, which does not address coastal area hazards, is presented in [3]. As an example, India has used systematically remote sensing (mainly Indian satellites) for many years to monitor coastal zones, gathering information regarding coastal habitats, landforms, shoreline, water quality, etc., and evaluating the impact of various natural hazards [4].
Using remote sensing tools for monitoring of the following five critical dynamic environmental phenomena is of major importance for decision-making regarding environmental risks for coastal and riparian areas: (a) land cover and land use, (b) coastline, (c) soil erosion, (d) land deformation, and (e) sea/ground water levels. In this work, the main objective is to design and develop an integrated monitoring system with decision support capabilities for coastal and riparian areas, based on periodically monitoring and extracting important knowledge regarding these phenomena. In the following, state of the art approaches that are currently used for monitoring these dynamic phenomena are briefly reviewed.
Land cover and land use (LCLU) of the Earth is changing dramatically because of human activities and natural disasters (e.g., deforestation, flooding, food shortage, greenhouse effect, unplanned urban extension, etc.). These environmental problems are often related to LCLU changes, so estimating LCLU maps and their changes is therefore very important for managing natural resources and monitoring environmental changes [5]. For this reason, many LCLU monitoring models have been developed for assessing change on land cover and use patterns for different purposes i.e., land cover and use changes have been monitored for instance in coastal zone management [6] and in wetlands management [7]. These maps are hard to produce and often have limited temporal resolution (e.g., Corine Land Cover maps [8] is available only for years 1990, 2000, 2006, 2012, 2018, while the European Space Agency (ESA) WorldCover map [9] is available only for the year 2020). Hence, the development of models that can estimate LCLU maps robustly from satellite data corresponding to specific time intervals is necessary for improved monitoring of LCLU changes.
Machine learning methods are typically used for handling these medium resolution satellite image time series. Early works on land cover classification utilized methods such as Random Forest [10,11], hidden Markov models [12,13], or Support Vector Machines [14], and were employed on hand-crafted features such as spectral statistics and phenological metrics [15,16]. Recent methods avoid the use of hand-crafted features and use Deep Neural Networks (DNNs), as they are capable of handling large amounts of data and creating more representative features. Convolutional networks have been adapted from the broader domain of computer vision to specifically create spatial representations of satellite image time series. Architectures, such as the popular U-Net model that was originally proposed for biomedical image segmentation [17], are able to create deep features from an input image and produce a meaningful segmentation that suites the task at hand. Specifically, the U-Net and its variations have been employed in remote sensing problems, producing accurate segmentation masks based on land cover/use [18,19,20,21,22]. The more recent Transformer architecture [23], first proposed to handle sequential data, has been adapted to create a new family of DNNs, the Vision Transformers (ViT), that seem to outperform convolutional methods in computer vision tasks [24,25,26,27].
Regarding LCLU classification, two state-of-the-art Deep Learning approaches were applied by combining data from multiple satellites. The former relies on Convolutional Neural Networks (CNN) and the latter on a hierarchical Transformer encoder. The accuracies of both methods, and all available band combinations, are compared using a very recent LCLU product as reference, namely ESA WorldCover.
Coastlines are the linking of land and ocean, which are strongly impacted by human activity [28]. Coastal and riparian zones host a rich variety of biodiversity and are valuable from an environmental, social, and economic point of view [29]. As technology progresses, EO assets, data, and technological tools provide a unique opportunity to establish methodological frameworks for the development of efficient decision support tools to monitor coastlines [30]. Through the years, various solutions have been proposed for severe coastline changes via soil erosions. These solutions usually exploit specific remote sensing indices such as NDWI, mNDWI, NDVI [31,32,33], etc., or more advanced classification approaches. By emphasizing the significant risks to which coastal zones are exposed, a decision support tool which can provide information on coastline status and trends through the years is a key asset for end users and policy makers.
Our proposed coastline monitoring module, offers the ability to provide past and current instances of the coastline status, including the export of multitemporal change maps that provide a better picture for the general coastal changes and threats. These outputs provide a tide-independent overview of a coastal area and can extract coast-line trends (water/land loss, soil erosion etc.) for better planning and recommended actions to be taken.
Furthermore, soil erosion, one of the most crucial environmental problems around the world [34], is directly related to land degradation, constituting its worst form, while causing severe environmental and socio-economic consequences [35]. According to a global estimate, during the mid-1990s, around one-third of the worlds’ cultivated land was affected by soil degradation, while about 56% of that land represented soil erosion [34,36]. Various studies [37,38] highlighted the severity of soil erosion, since about 80% of the total cultivated lands indicate a moderate to severe erosion rate, which resulted in the loss of approximately 10 million hectares of arable land worldwide [34,39]. In this work, a popular empirical model, namely RUSLE (Revised Universal Soil Loss Erosion, [40]), is applied for calculating the average seasonal soil erosion.
Riverine, riparian and river delta zones are usually affected from land subsidence and surface deformation events due to severe soil erosion and groundwater withdrawal phenomena. To this end, Synthetic Aperture Radar (SAR) data have successfully generated large scale land subsidence and surface deformation maps through InSAR (Interferometric Synthetic Aperture Radar) and DinSAR (Differential InSAR) techniques [41]. The use and integration of freely available multi-temporal SAR data (e.g., Copernicus Sentinel-1) into such methodologies ensures the liveability and the end-users’ acceptance of the end products.
In order to assess both groundwater as well as sea level variations in the study area, in situ sensors were deployed. Real-time sea level measurements are derived from a tide gauge station, while measurements of the water levels of specific boreholes are obtained (either by direct measurement or using previously installed piezometer sensors) and used to create piezometric maps.
In this context, to offer better services to end users (civil protection authorities, citizens, etc.), data from all above sources were used to periodically extract new knowledge regarding different coastal zone hazards, assess associated risks, produce early warnings, develop accurate prediction models based on past data, and assist users in decision-making for mitigating these risks. Decision support tools are computer-based tools designed to assist in decision-making, including the management of environmental risks and long-term planning for impacts in different sectors and regions. A recent extensive review of previous work on Decision support tools and systems for coastal planning and management is made in [42]. In this work, we opt for a simple rule-based Decision Support System (DSS) that will be developed to evaluate changes over time and create alerts as needed. The rules for creating alerts will then be validated and fine-tuned together with the participating end-users, to ensure successful end-user adoption of the proposed system.
The extracted knowledge from the aforementioned modules, including links to files stored in a File Repository, is stored to a MySQL database in the central server. In addition, a WebGIS interface is developed as a Progressive Web Application, allowing the user to query and visualize data from the server and/or external sources, providing rich information for decision support, such as high-level monitoring information, recommendations, alerts (when needed), etc. This research is conducted within the framework of “ExPert Integrated support system for CoastaL mixed urbAn—industrial—critical infrastructure monitoring using Combined technologies” (EPIPELAGIC) co-financed from the European Union, China, and Greece.
The main contributions of this work can be summarized as follows:
  • Two state-of-the-art Deep Learning approaches are applied for LCLU classification using input and products from multiple satellites. The former relies on CNN and the latter on a hierarchical Transformer encoder. The accuracies of both methods, and all available band combinations, are compared, using a very recent LCLU product, namely ESA WorldCover, as reference.
  • A simplified methodology utilizing EO multispectral datasets to produce a periodically tide independent status of a river delta coastline.
  • Calculation of the average seasonal soil erosion in the Aliakmon and Axios river catchments in Thermaikos Gulf is performed using the RUSLE model. The soil erosion in the catchments has a considerable impact on the studied coastal area.
  • Robust generation of large-scale land subsidence and surface deformation maps based on freely available multi-temporal SAR data.
  • An integrated monitoring system, collecting and analysing information regarding natural hazards, which uses a DSS to assess changes over time and produce alerts when needed and an online WebGIS application to access and present this information to the end users.

2. Materials and Methods

2.1. Study Area

The study area is shown in Figure 1 and is an area of approx. 762 km2. It concerns the coastal area of Thermaikos Gulf and includes part of the city of Thessaloniki and the Axios Delta National Park (Axios National Park—Loudia—Aliakmon). This park is of high ecological importance and its protection has been included in the Ramsar Convention as a Wetland of International Importance and in the network of Natura 2000 sites. Additionally, it has been included in the Important Bird Areas (IBA). This area was chosen to be studied under EPIPELAGIC because it is a coastal area close to a river delta and combines a variety of land uses (urban, industrial, commercial zone) and a variety of economic activities, which in many cases are in conflict with environmental protection and have significant environmental impacts. It is therefore an optimal choice for the study area under EPIPELAGIC. Furthermore, since the soil erosion observed in the catchment areas of the aforementioned rivers (Figure 1) has an impact on the study area, the assessment of the average seasonal erosion rate per unit area of these catchments was carried out accordingly.
The pilot study will use periodic cloud-free composites from a number of satellites between 1990 and 2020, including Sentinel-2 (Level 2A), Sentinel-1 Level-1 Ground Range Detected (GRD) products, Landsat-5, Landsat-8, and ERS-1/2 missions. In addition, data and infrastructure used in previous studies/research projects (e.g., INDES MUSA project [43]) will also be utilized, including any existing digital databases of the study area. In-situ measurements at the study area are also available for: (a) measuring the displacement of at least 15 points on the ground of the study area, with the aim of recording soil movement within one year, with an accuracy of less than one centimetre, (b) monitoring sea level changes using measurements obtained from an already installed tide gauge, and (c) monitoring groundwater pressure variations using a set of piezometer sensors that were already installed.

2.2. Platform Architecture

The basic architecture of the proposed platform is illustrated in Figure 2 and consists of four main subsystems: (a) a data collection and processing module, (b) five knowledge extraction modules, (c) the DSS, and (d) the WebGIS User Interface. These subsystems are described in more detail in the following sub-sections.

2.2.1. Data Collection and Pre-Processing

A robust methodology has been implemented to periodically collect and pre-process data (to reduce noise and distortions) from several heterogeneous data sources, including satellite data (Optical/Multispectral, SAR, High Resolution images, Digital Elevation Model(s) (DEMs), etc.) and in-situ real-time measurements (tide gauge, Global Positioning System/Global Navigation Satellite System (GPS/GNSS) etc.). For more efficient data access and computational efficiency, the Google Earth Engine (GEE) [44] cloud platform for satellite data pre-processing was also used. Specifically, GEE was used to create composites from Sentinel 1, Sentinel 2, Landsat 5, Landsat 8 imagery for the study area within pre-specified 6-month time periods (April-October/October-March) for the years 1990–2020. These optical and SAR composites are stored into the main database, along with other satellite products that are not currently available in GEE (e.g., ERS or Sentinel-1 SLC data). The latter need to be individually downloaded and pre-processed, e.g., using the SNAP software [45] of the ESA to create similar composites. More specifically, GEE provides a very convenient and efficient way to pre-process all available images in the cloud using simple JavaScript or Python scripts. For multispectral optical images, namely ESA Sentinel 2 (Level 2A) [46], NASA Landsat 5, and Landsat 8 [47] products, a common pre-processing procedure was used: First, all available images within the specified 6-month time range are clipped to the Area of Interest and all bands are interpolated into 10 m/px resolution. For these imagery, cloud, cirrus, and shadow masks are available and mapped over each individual image in order to create an image composite with 12, 6, and 7 multi-spectral bands, respectively. To produce these images, temporal median filtering on all available cloud-cirrus and shadow free images is applied within the selected time period. On the other hand, Sentinel-1 [48] SAR data (Level-1 Ground Range Detected (GRD) product) is also available in GEE and is already pre-processed via ESA SNAP/Sentinel-1 Toolbox. The pre-processing steps are Thermal Noise Removal, Radiometric Calibration, and Terrain Correction. The pre-processing procedure is similar to that used for optical images, however, no cloud masking is required, as SAR signals are not absorbed by clouds. Hence, a SAR composite with 2 bands (composites for backscattering signals for VV and VH dual polarization) is finally created by temporal median filtering of all available images within the selected area of interest and time period. Again, both bands are interpolated into 10 m/px resolution so as to facilitate fusion with optical data.
Regarding ERS data or Sentinel 1 Single Look Complex (SLC) SAR data, containing phase information, which is necessary for SAR Interferometry, these are not available in GEE, so local pre-processing was used. Specifically, after downloading the images from the ESA repository, a sequence of corrections and transformations for fusion preparation is recommended in the literature [49]. First, the image is clipped to the Area of Interest. Updated orbit state vectors are then uploaded and applied to the data. This information will then be used in subsequent editing steps to improve the quality of the result. This is followed by the removal of thermal noise. This correction concerns the filtering of additional noise in the signal to improve the result, especially in places with low reflection. Then, a correction is made for noise that occurs mainly at the edges of the image (border noise removal). This removes low-power noise and invalid data at the edges of the image. Then, the signal is calibrated, where the values of the reflection coefficient σ0 (backscatter coefficient) are extracted. Another filter is applied to the generated signal, which removes the speckle filtering noise. Last but not least it is followed by a terrain correction procedure, which corrects geometric deviations of the pixel values towards their correct image locations. Additional useful auxiliary data are also available from GEE, namely:
  • Corine Land Cover (CLC) Layers for different years (e.g., 1990, 2000, 2006, 2012, 2018) can be used as ground truth for training new Land Cover classification approaches. In EPIPELAGIC, we are mainly interested in Level-1 CLC labels (i.e., Artificial Surfaces, Agricultural Areas, Forests and seminatural areas, Wetlands, Water bodies).
  • The WorldCover v100 Layer for the year 2020 can also be used as a reference for training and evaluations Land Cover classifications. WorldCover has an improved 10 m/px resolution, while CLC has a 100 m/px resolution. However, a disadvantage of WorldCover is the small number of classes, namely 11 classes (and only 8 are available in the specific study area), while CLC 3-level hierarchical classification system has 44 classes at the third (most detailed) level.
  • NASA SRTM Digital Elevation 30 m Model will be used for (a) calculating soil erosion maps and (b) calculating flood risk maps due to sea level rise.
  • JAXA ALOS World 3D—30 m is a global Digital Surface Model (DSM) which will be used for Land Cover Maps generation with Deep Neural Networks.
All pre-processed EO/remote sensing products, as well as the available in-situ measurements (GPS/GNSS, tide-gauge and piezometric measurements), are stored in a local file repository, following specific naming conventions. Furthermore, the database stores references to data files (full path names) and other relevant metadata (e.g., data source. capture date, etc.) to facilitate data access and browsing.

2.2.2. Land Cover Land Use Classification

To provide reliable information to the change monitoring and DSS, accurate and regularly updated land cover maps are needed. As the availability of such land cover maps is typically very low, we leverage Deep Learning algorithms to tackle the complicated and time-consuming task of remote sensing images classification. More particularly, two semantic segmentation algorithms were employed, namely (a) the CNN-based U-Net algorithm [17] and (b) the SegFormer, i.e., a more modern Transformer-based algorithm [27]. These algorithms were trained on a PC with NVIDIA RTX 3080 GPU equipped with 12 GB memory. The more important python libraries used are PyTorch, NumPy, and Rasterio. Both algorithms were trained on 100 epochs, 12 batch size, 10−4 learning rate, and the Adam optimizer. As for the loss function, first the Cross-Entropy loss function for both algorithms was used. As the dataset is highly imbalanced (more than 75% of the dataset consists of four out of eight classes), the Focal Loss function was also used as a more effective alternative for dealing with class imbalance [50]. Focal Loss is a dynamically scaled cross entropy loss where the scaling factor decays to zero as confidence in the correct class increases. Through trial and error, the best accuracy was achieved with gamma (γ) set to one (1).
Figure 3 illustrates the machine-learning pipeline where remote sensing data from different satellites, regions, and timelines are downloaded and pre-processed in GEE. More specifically, Sentinel-2 (L2A), Sentinel-1 (GRD, Landsat-8, Sentinel-2 indices (NDVI, NDBI, NDWI) [32,33,51], and DEM/SLOPE maps from the Japanese Aerospace Exploration Agency were downloaded. These imageries contain a total of 22 Bands, while the two open available Land Cover maps from the ESA, CLC, and WorldCover products are used as target data. Initially, CLC was planned to be used for this module, but given the recent release of the significantly improved (both in terms of resolution and accuracy) WorldCover product, the CLC layer was not utilized.
As the main study area, namely Thermaikos Gulf, is a coastal and riparian zone, the training set consists of various landscapes throughout Greece, with similar characteristics [52]. In order to use more coastland land cover samples, inland regions that are close to water areas (rivers or lakes) were selected, shown in Figure 4 within the yellow polygons. Similarly, the main study area used only for testing is depicted within the red polygon in Figure 4.
As the different regions used to obtain the Training and Testing sets have different sizes, all multispectral images were split into tiles, each containing 256 × 256 pixels. After splitting these regions into 256 × 256 sized patches, a dataset consisting of 4258 Training patches and 204 Validation and Testing patches was created.

2.2.3. Coastline Monitoring

For the scope of monitoring the coastal environment in the study area, a continuous flow of satellite data and information needs to be available. To this end, Multispectral timeseries data from Landsat and Sentinel missions will be exploited. Overall, for the data acquisition process the selection criteria of <5% cloud cover was set, and a quality control task was performed for no-data stripes and/or corrupted scenes. In this work, only available data from Landsat 5 (2000–2011), Landsat 7 (2000–2004), and Landsat 8 (2013–present) were used, while Sentinel-2-data will also be used in the final version. An additional crucial data handling step was to enhance the spectral capacity of the raw data through the implementation of new datacube composites by adding on top of the existing spectral bands additional selected indices (NDWI, MNDWI, and NDVI), acknowledging the fact that coastline identification is a challenging task by itself.
A synoptic workflow of the developed methodology for the creation of mean maps, depicting the coastline state at each period, is presented in Figure 5.
All implemented technical steps of this module have been executed through a series of Python-developed functions in a unified script. To this end, Python libraries such as GDAL, OpenCV, rasterio, sklearn, pytorch, pandas, etc. were widely utilized in this study. As typically implemented in most pixel-based multispectral data handling process, we did not treat each corresponding spectral band separately but on the other hand proceeded with the creation of easy to manipulate and comprehend spectral datacubes for each scene, while the initial step prior to atmospheric correction was to crop all datasets with a basic Python function. As a final step, an unsupervised classification through k-means algorithm is executed to each generated datacube for the production of the classification maps.
The extraction of yearly average coastline status and the flood probability maps is generated after the analysis each consequent image product produced from each observation, as presented in the results in a following section.

2.2.4. Soil Erosion Estimation Using the RUSLE Model

Coastal areas, as already mentioned, are vulnerable to anthropogenic pressures, leading to diverse problems that might result in the degradation of their environmental status. Among them is soil erosion that could have significant environmental and economic impacts; its assessment is considered, therefore, of major importance [35]. In this paper, the Revised Universal Soil Loss Equation (RUSLE) [40] is used to assess the average seasonal soil erosion in the Aliakmon and Axios river catchments in Thermaikos Gulf, Northern Greece; areas of important environmental and economic value with a considerable impact on the study area (Figure 1).
The RUSLE model constitutes the amended version of the Universal Soil Loss Equation (USLE) empirical model, initially presented by [53], which was improved and further developed, through the embodiment of supplementary data and the incorporation of relative research results [54]. The model was developed by the USDA (United States Department of Agriculture) and is a standard method to estimate soil loss by predicting the average soil loss over time [40]. The model enables the comparison of methods and results and can be implemented internationally since many countries could adopt and apply it [55]. Combined with GIS and remote sensing, RUSLE can be applied for the average soil erosion estimation in terms of spatial distribution and sediment deposition [56]. In this paper, the RUSLE model was applied twice for each catchment area; for winter (November–April) and summer (May–October). The average seasonal gross erosion was estimated using the following equation developed by [57]:
A = R × K × L S × C × P
A: the average seasonal soil loss selected for specific R
R: the rainfall-runoff erosivity factor
K: the soil erodibility factor
LS: the slope length and steepness factor
C: the cover management factor and
P: the conservation practice factor
The necessary data for the calculation of each factor and the final application of the RUSLE model are presented in the flowchart below (Figure 6). All the datasets were stored in two (2) geodatabases—for winter and summer—in raster format with a pixel size of 250 m with the same coordinate system, and all the necessary calculations were implemented in a GIS environment using the ESRI 2020, ArcGIS Desktop: Release 10.8 software (Redlands, CA: Environmental Systems Research Institute).
The calculation of the Rainfall Erosivity Factor (R) was based on satellite data, and more specifically a 20-year monthly rainfall dataset (from 2000 to 2020) was used to obtain the mean seasonal rainfall. Soil data were used in this study to calculate the K-factor, which was considered to be stable all over the year, while the calculation of the LS-factor was conducted using the DEM. The specific factor is contingent on both the slope length factor (L) and the slope steepness factor (S) and is responsible for the topographical effects on soil erosion. The C-factor also has a significant effect on soil erosion, and according to [58] is related to the NDVI index, as such indices are considered to be of major importance for the extraction of vegetation parameters. Finally, P-factor was based on the land cover type, displaying how LCLU have an impact on soil erosion, and it was determined through the consideration of an extensive literature review [59], where each land use class was correlated with a coefficient value ranging from 0 to 1. The aforementioned data used for the application of the RUSLE model and the relevant sources are presented in Figure 7. It should be noted here that the LCLU classification calculated in this work concerns a limited area in the coastal zone of Thermaikos Gulf including part of the city of Thessaloniki (Figure 1). However, in order to assess the average soil erosion in the relevant catchments, the LCLU was needed for a quite extended area, as shown in Figure 1. Therefore, the CLC was used, which is a fully satisfactory land cover classification for the calculation of the P-factor in the RUSLE model.

2.2.5. Land Deformation

SAR data usage has gained ground over the past decades in several applications due to their evident advantages over other EO datasets like optical multispectral imagery. Such technical advantages include the capability to identify small land movement and land subsidence changes and the capability of Microwave signals to penetrate clouds and in general to provide weather-independent reliable data. EPIPELAGIC project exploits SAR by the utilization of ERS and Sentinel 1 data from the ESA, through InSAR.
On top of the aforementioned InSAR application, GPS/GNSS from in-situ surveys (Figure 8) and permanent stations are deployed for this task. Specifically, two separate GNSS surveys at the wider case-study area have taken place on 2021 and 2022, while permanent GNSS stations are constantly feeding into the server with real-time information. This rich information is used under the project’s framework, both on top of InSAR results as validation/verification ground truth and as an external source of information that contributes directly to the DSS.

2.2.6. Sea/Ground Water Level

Sea level rise is currently accelerated by the climate change and may have significant consequences for coastal societies. Detecting, attributing, and understanding the contemporary changes and trends in sea level has great importance in constraining projections of future sea level and in preparing the adaptation of coastal communities [65]. Although sea level measurements can also be accurately obtained using satellite altimetry, in this work, we rely on real-time in-situ measurements using a tide gauge station (Figure 9). Furthermore, measurements of the groundwater levels at specific boreholes are periodically obtained (either by direct measurement or using previously installed piezometer sensors) and used to create piezometric maps.

2.3. Decision Support System and WebGIS Interface

A Decision Support System (DSS) is under development that will allow the end-user to analyse and evaluate the available data from the aforementioned subsystems and receive alerts and recommendations, as required. It will detect changes in the data periodically produced by the knowledge extraction modules, assess their importance, evaluate risk factors, and issue alerts/notifications to the end user (when required). More specifically, it will monitor noticeable changes in land cover or use, e.g., urban expansion and loss of vegetation, coastline changes, soil erosion, land deformation, and sea/ground water levels, and create corresponding change maps. Results produced by the DSS that can be used to assist the decision-making process (e.g., change maps, risk probability maps, simulation results, etc.) will be stored back to the database to be made available to the end-users. The DSS will also be able to issue alerts that are communicated to the end-users via emails or push notifications.
Finally, a WebGIS is under development as the frontend (User Interface) of the system. It was developed as a Progressive Web App (PWA), combining merits of modern internet browsers and mobile apps, i.e., progressive display in different systems, cross platform support, use of standard web development technologies, and ability to run offline. The WebGIS interface allows the user: (a) to query and visualize raster or vector data from the database, as well as data from external sources or user-uploaded data (e.g., in situ measurements), (b) to create various informative graphs, e.g., plot of mean risk factors (e.g., urban areas, erosion, land deformation, sea/ground water levels) in the study area with respect to time, (c) to create flood map simulations indicating the areas that will get flooded, assuming a certain sea water level.
The WebGIS interface supports the display of multiple raster and vector layers. A widget layer is provided that allows the user to select specific layers for display and alter viewing options (e.g., transparency) for each layer.
All data exchange between the server backend, namely the file repository and database, and the web app frontend, are realized via a REST API that is implemented using python, based on flask micro web framework and MySQL-connector-python for MySQL access. Its main tasks are (a) to retrieve images, in situ-data and information stored in the database, including high-level monitoring information and DSS data/recommendations, (b) to provide alerts, as needed, via emails or push notifications to the web app, (c) to store specific in-situ information provided by the user, e.g., GPS/GNSS measurements, to the database.

3. Results

3.1. Land Cover Land Use Classification

A comparison between the two Deep Learning models used in this paper, namely U-Net and the SegFormer, will be presented. In order to optimize the performance of the two approaches, different band combinations from various satellite data were used as input. Due to limited space, we will discuss only the combinations providing the highest accuracies. Two accuracy metrics, namely average (avg) F1-Score (1) and Overall Accuracy, OA, (2) were used.
O A = T P + T N T P + T N + F P + F N
a v g F 1 = 1 C 1 C 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l
C: Number of Classes
TP: True Positive
TN: True Negative
FP: False Positive
FN: False Negative
Precision: TP/(TP + FP)
Recall: TP/(TP + FN)
Specifically, the best band combination for the U-Net algorithm uses 19 input channels: all 12 Sentinel-2 Bands, three spectral indices (NDVI, NDBI, NDWI), two Sentinel-1 GRD bands (VV and VH), and DEM/Slope products. Regarding the loss function, Focal Loss was used.
For the SegFormer, the best band combination consists of 13 input channels: six Sentinel-2 bands (B2, B3, B4, B8, B11, B12) as well as three spectral indices (NDVI, NDBI, NDWI), two Sentinel-1 GRD bands (VV and VH), and DEM/Slope products. In this case the Cross-Entropy loss function was applied.
The generated land cover maps from the two models for the study area are shown in Figure 10 and Figure 11, respectively. For both cases, the WorldCover Layer (Figure 12), was used as ground truth data. The different colour-coding used is illustrated in Table 1.
Both algorithms provide very satisfactory results. The Overall Accuracy (OA) is 65% and 64% for U-NET and SegFormer, respectively, while the avgF1-Score is 42.125% and 49.75%, respectively.

3.2. Coastline Monitoring

Based on the workflow presented in Section 2.2.3 for the module of coastline monitoring, two different set of products can be extracted:
  • Yearly Average probability maps
  • Change detection maps
As the validation of the outputs and fine-tuning of algorithmic in the study area is not completed yet, the module was tested in a smaller case study (Evros River delta—NE Greece) with similar characteristics and some indicative results are presented below:

3.2.1. Yearly Average Probability Maps

In Figure 13, various generated yearly average maps (for the years 2000, 2005, 2010, 2015, and 2020) are provided, as produced from our coastline monitoring module pipeline.

3.2.2. Temporal Change Maps

Change maps over a selected time window can be generated by analysing and aggregating the yearly outputs. Consequently, through this analysis, the temporal change maps for the selected periods have been created, revealing the more pressured areas of (i) Receding water bodies and (ii) New water bodies. Below, in Figure 14, these change maps are demonstrated accordingly for a respective time window of 5 years.

3.3. RUSLE

The outputs of the RUSLE model, displayed in Figure 15 and Figure 16, indicate the potential seasonal soil erosion for the Aliakmon and Axios river catchments. The high-risk erosion areas concerning the Axios catchment, both for winter and summer, are mainly observed at the north and west parts of the catchment, as well as at some areas of lower extent at the southern and north-eastern outer limits. The erosion, however, is more intense during the winter, which may be due to the heavy rainfall with a longer duration. Finally, the areas with lower soil loss are observed near the catchment’s outlet, and in low altitude areas. In the Aliakmon catchment, for both seasons, the high-risk erosion areas are mainly observed in the central areas of the catchment and those with high relief. During winter, these areas are more extended compared to summer and high-risk erosion areas are also located at the north and west parts of the catchment, as well as at its south-western boundaries. The lower soil loss is observed in the lowlands of the catchment and near the outlet.

3.4. Land Deformation

For the identification of long-term land subsidence changes over the study area, InSAR approaches are implemented over different periods of time. Space-borne radar interferometry is a technique that can measure ground movement between two radar images acquired at different times on the same area on a pixel-by-pixel basis [66]. Interferometry allows the monitoring of even slight ground movement, down to a few millimetres, across wide areas.
To this end, in EPIPELAGIC an interval period of 2 years is selected, and the respective velocity change maps are generated. The two-year interval is considered sufficient for the identification of the long-term changes over the wider study area. The methodology follows all necessary SAR data pre-processing steps and the generation of the change detection velocity maps through SNAP software functionalities, deployed from python calls and snappy library. The overall methodology and processing workflow follows the tutorial (Serco Italia SPA, 2018) provided by the Research and User Support for Sentinel Core Products initiative. A result of such velocity map for the changes from the years 2014 to 2016 can be seen below in Figure 17.
For the identification of micro-movements and SAR deriving results validation purposes, GPS/GNSS surveys are implemented, and the measurements record of an already established GNSS network is being analysed. To this end, 2 GNSS surveys have been performed over 19 selected ground points (Figure 18), including critical infrastructure and areas presenting land cover changes over the years.

3.5. Sea/Ground Water Level Measurements

A tide gauge automatic measuring station of the sea tide has been installed at the southeast side of the commercial port of Thessaloniki for the EPIPELAGIC project. Sea level measurements obtained from this tide gauge are communicated in real time to the server and can also be used to create plots for specific time periods upon request (Figure 19).
Furthermore, in order to monitor ground water variations, two surveys have been planned, within the summer of 2022 and 2023, respectively, in order to create piezometric maps for the study area, after collecting measurements from a set of boreholes in the area managed by the Thessaloniki Water Supply and Sewerage Company.

3.6. WebGIS

In Figure 20a–c, three screenshots for the current (preliminary) version of the WebGIS Interface are illustrated: the User Menu, a RGB image composite from Sentinel-2, and the urban area change map between years 1990 and 2018. The layer widget for selecting viewing options for background/foreground layers is also shown in Figure 20c.

4. Discussion

In this paper, an integrated system is presented, designed for monitoring of coastal and riparian areas for natural and anthropogenic hazards. The system leverages state of the art technologies of remote sensing data and machine learning for monitoring several phenomena that may strongly affect these areas, especially given climate change: (a) Land Cover Land Use, (b) Coastline, (c) Soil Erosion, (d) Land deformation, and (e) Sea/ground water levels. For this reason, five separate knowledge extraction modules are designed and implemented, combining information from medium resolution satellite images, external data sources, and in-situ measurements. The extracted information is stored to a central database and a rule-based Decision Support System is developed to monitor changes over time, create alerts when required, and provide insights to end-users. Finally, a WebGIS interface was also designed and developed, allowing end-users to query and visualize data from the database, receive alerts (when needed), etc.
Regarding LCLU estimation, two state-of-the-art deep neural network approaches were used, namely the U-Net network [17], based on Convolutional Neural Networks (CNN), and the SegFormer network [27], a Transformer-based model. Both networks are widely used in segmentation tasks using RGB images, so some modifications were needed to accommodate multiple input channels, exploiting data from different satellite images and products. Although U-Net and Transformers have already been used for Land Cover and Land Use (LCLU) classification from satellite images [67,68,69,70,71,72], our approach is one of the few that attempts to identify the combination of bands from different freely available satellite images and products so as to optimize the training accuracy.
Concerning coastline monitoring, our approach exploits all available information from multiple multispectral satellites (Landsat, Sentinel-2) and can create multitemporal change maps that can better visualize the coastal changes and threats. Thus, a tide-independent overview of the AOI is obtained and coast-line trends (water/land loss, soil erosion, etc.) can be identified.
In respect of the soil erosion in the area, results obtained indicated that the RUSLE model can be successfully applied for this purpose in Mediterranean catchments, even in areas where the extremely varying flow regime throughout the seasons and the lack of available data constitute a challenge [73,74,75]. Soil erosion is a natural process that changes over time, since it is greatly influenced by conditioning factors, such as rainfall and vegetation. Remotely sensed rainfall and vegetation data formed the basis for creating the suitable dynamic rainfall-runoff erosivity (R) and cover management (C) factors for the application of such an empirical model as RUSLE, while the remaining three factors were maintained stable, due to their static “nature”. RUSLE model can offer an explicit prospect for understanding the interaction among soil erosion and relevant factors, since it is a robust, cost-effective, less data demanding, and less time-consuming approach.
Regarding land deformation, InSAR and DinSAR techniques [41] using Synthetic Aperture Radar (SAR) data from Copernicus Sentinel-1 Single Look Complex (SLC) products were found to be very successful in generating large scale land subsidence and surface deformation maps.
Online monitoring of both ground water levels using in-situ measurements is very important for coastal communities, e.g., to evaluate changes in the resource/tide over time, to develop ground-water/tide models and forecast trends, to control traffic flow in ports with shallow waters, etc.
The WebGIS and DSS modules were both designed with an intention to maximize usability and the user experience. For this reason, a very simple rule-based DSS will be provided, which can be easily fine-tuned according to end user preferences. Similarly, the WebGIS is implemented as a Progressive Web Application, exploiting the advanced capabilities of modern internet browsers, simplifying development and future updates, offering cross platform support and improved user experience.
However, the system also has some limitations. For instance, application of the proposed LCLU estimation techniques in large areas may require high memory and computational resources. Specifically, as the target resolution is 10 m, the input satellite images need to be split to smaller tiles both for local or cloud pre-processing, as well as for applying the proposed deep learning models. Another problem is that the training of these models in this work used mostly coastal areas, so new-training nay be required to improve accuracy for non-coastal (e.g., mountainous) areas. This also limits the potential use of the proposed LCLU estimates within the RUSLE model.
Using forecasting for both coastline monitoring and LCLU was also considered, and some preliminary tests using LSTM (Long-short term memory) [76] recursive neural networks with time series of satellite data were performed. However, such forecasts rely on a small set of previous observations, so they may underestimate the acceleration of climate change’s effects.
A first version of the proposed monitoring system is currently available and evaluated by end users from the managing authority of the study area, however the final version of the system will be finalised within 2023. In parallel, the team from China that participates in the EPIPELAGIC project is currently developing systems for assessing inundation and erosion phenomena Gudong Seawall in the Yellow River Delta [77]. Future extensions of the proposed system may integrate such additional monitoring systems, thus further increasing the impact of the system towards more application scenarios in China and worldwide. The easy-to-configure-and-use design concept of the proposed system will contribute significantly towards this aim.
The characteristics of the technical means of seawall deformation monitoring are analysed from three perspectives: space, ground and underwater. The advantages and disadvantages of each technical means are compared also. According to the design requirements of the seawall deformation monitoring system, combined with the harmful factors of the seawall deformation, a seawall deformation monitoring program that integrates space, ground, and underwater is proposed, which provides a new idea for the seawall deformation monitoring work. This program provides an in-depth analysis of monitoring points, system design, and selection of technical means from three levels: long-term continuous monitoring, periodic monitoring, and instant monitoring.

5. Conclusions

EPIPELAGIC is a project that will suggest solutions and tangible measures for mitigating the effects of climate change in coastal areas as well as adapting them for use in a wide range of data. The main objective is to create a platform based on the latest remote sensing and machine learning technologies for monitoring and evaluating important changes that can be associated with natural and anthropogenic hazards, including LCLU and coastline changes, soil erosion, land deformation, and sea/ground water levels.
By feeding freely available medium resolution satellite data to robust deep learning algorithms, we can achieve accurate LCLU estimation with an increased temporal resolution. Furthermore, using the same data, we can robustly estimate a periodically tide independent status of a river delta coastline. Despite the fact that the RUSLE model is a relatively simplistic model for the estimation of soil erosion, it is one of the most widely used models, quite simple in parametrization and with low execution time requirements. Based on the mapping of areas that are vulnerable to erosion in the relevant catchment areas, decision-making could be further supported in a timely and successful manner regarding the adoption of measures to manage erosion in order to limit its impact on the nearby coastal zone. Using radar techniques with freely available multi-temporal SAR data, we can produce large scale land subsidence and surface deformation maps. Finally, the platform can also collect and process in-situ measurements, such as sea and ground water levels.
The proposed platform, with the on-line monitoring and assessment of the impact of climate change, will have a major impact on financial challenges related to the priorities of investors and investment banks, insurance companies and the value of the properties themselves. Regarding the social dimension of the problem related to the safety of citizens, through the platform it will be possible to support a better and more sustainable planning of areas on a local scale (location of protection projects, development projects, etc.), leading to safer living of citizens in coastal areas. From an environmental point of view, the project will implement a risk analysis, identifying the vulnerability of areas related to hazards such as coastal erosion, soil erosion, floods, etc., with a view to their sustainable management and rapid response. EPIPELAGIC will support efficient and faster risk analysis and management using high-resolution models and large-scale data processing to provide a fast and reliable Solution as a Service (SolaaS). This will increase the time available for the coordination and organization of emergency measures. All alarm warnings will be addressed to end users (decision makers, municipal authorities and relevant ministries, industrial and insurance companies, etc.) for better cooperation and coordination. Finally, the project supports the development of new research activities and innovative tools while at the same time strengthening the cooperation between the academic and industrial sector, as well as the transnational cooperation between Greece and China.

Author Contributions

Conceptualization, V.C., D.K.; methodology, A.T., N.G., C.K. and A.P.; software, A.T., N.G., C.K. and A.P.; validation, A.T., N.G., C.K. and A.P.; data analysis, A.T., N.G., C.K. and A.P.; writing—original draft preparation, A.T., N.G., C.K. and Z.P.; writing—review and editing, T.N. and D.K.; supervision, D.K.; project administration, V.C.; funding acquisition, V.C. All authors have read and agreed to the published version of the manuscript.


This research was funded by European Union and Greece, grant number T7ΔΚI-00160.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Not applicable yet, but link to the current version of the platform will be given in the manuscript is accepted.


This work was based on procedures and tasks implemented within the project “ExPert Integrated support system for CoastaL mixed urbAn—industrial—critical infrastructure monitoring using Combined technologies—EPIPELAGIC”, which is co-financed from the European Union and Greece under the operational programme «Competitiveness, Entrepreneurship, Innovation» in the announcement “Bilateral and Multilateral S&T Cooperation Greece—China” (development code: T7ΔΚI-00160).

Conflicts of Interest

The authors declare no conflict of interest.


  1. Kosta, A.; Paraskevopoulos, I.; Agapiou, A.; Battistin, F.; Serpetti, M.; Waldoch, F.; Hadjimitsis, D. Remote sensing techniques for archaeology: A state of art analysis of SAR methods for land movement. In Proceedings of the SPIE Eighth International Con-ference on Remote Sensing and Geoinformation of the Environment (RSCy2020), Paphos, Cyprus, 16–18 March 2020; Volume 11524, pp. 105–119. [Google Scholar]
  2. Coastal Zones, Copernicus Land Monitoring Service. Available online: (accessed on 16 August 2022).
  3. McCarthy, M.J.; Colna, K.E.; El-Mezayen, M.M.; Laureano-Rosario, A.E.; Méndez-Lázaro, P.; Otis, D.B.; Muller-Karger, F.E. Satellite remote sensing for coastal management: A review of successful applications. Environ. Manag. 2017, 60, 323–339. [Google Scholar] [CrossRef] [PubMed]
  4. Nayak, S. Coastal zone management in India− present status and future needs. Geo-Spat. Inf. Sci. 2017, 20, 174–183. [Google Scholar] [CrossRef]
  5. Muttitanon, W.; Tripathi, N.K. Land use/land cover changes in the coastal zone of Ban Don Bay, Thailand using Landsat 5 TM data. Int. J. Remote Sens. 2005, 26, 2311–2323. [Google Scholar] [CrossRef]
  6. Tassi, A.; Vizzari, M. Object-oriented lulc classification in google earth engine combining snic, glcm, and machine learning algorithms. Remote Sens. 2020, 12, 3776. [Google Scholar] [CrossRef]
  7. Faruque, M.J.; Vekerdy, Z.; Hasan, M.Y.; Islam, K.Z.; Young, B.; Ahmed, M.T.; Kundu, P. Monitoring of land use and land cover changes by using remote sensing and GIS techniques at human-induced mangrove forests areas in Bangladesh. Remote Sens. Appl. Soc. Environ. 2022, 25, 100699. [Google Scholar] [CrossRef]
  8. CORINE Land Cover. Available online: (accessed on 29 August 2022).
  9. Zanaga, D.; Van De Kerchove, R.; De Keersmaecker, W.; Souverijns, N.; Brockmann, C.; Quast, R.; Arino, O. ESA WorldCover 10 m 2020 v100; Zenodo: Geneve, Switzerland, 2021. [Google Scholar]
  10. Inglada, J.; Arias, M.; Tardy, B.; Hagolle, O.; Valero, S.; Morin, D.; Dedieu, G.; Sepulcre, G.; Bontemps, S.; Defourny, P. Assessment of an operational system for crop type map production using high temporal and spatial resolution satellite optical imagery. Remote Sens. 2015, 7, 12356–12379. [Google Scholar] [CrossRef]
  11. Pelletier, C.; Valero, S.; Inglada, J.; Champion, N.; Dedieu, G. Assessing the robustness of Random Forests to map land cover with high resolution satellite image time series over large areas. Remote Sens. Environ. 2016, 187, 156–168. [Google Scholar]
  12. Siachalou, S.; Mallinis, G.; Tsakiri-Strati, M. A hidden Markov models approach for crop classification: Linking crop phenology to time series of multi-sensor remote sensing data. Remote Sens. 2015, 7, 3633–3650. [Google Scholar] [CrossRef]
  13. Giordano, S.; Bailly, S.; Landrieu, L.; Chehata, N. Improved crop classification with rotation knowledge using Sentinel-1 and -2 time series. Photogramm. Eng. Remote Sens. 2020, 86, 431–441. [Google Scholar] [CrossRef]
  14. Devadas, R.; Denham, R.J.; Pringle, M. Support vector machine classification of object-based data for crop map-ping, using multi-temporal landsat imagery. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2012, 39, 185–190. [Google Scholar] [CrossRef]
  15. Qiong, H.U.; Wu, W.B.; Qian, S.O.N.G.; Miao, L.U.; Di, C.H.E.N.; Yu, Q.Y.; Tang, H.J. How do temporal and spectral features matter in crop classification in Heilongjiang Province, China? J. Integr. Agric. 2017, 16, 324–336. [Google Scholar]
  16. Nguyen, L.H.; Joshi, D.R.; Clay, D.E.; Henebry, G.M. Characterizing land cover/land use from multiple years of Landsat and MODIS time series: A novel approach using land surface phenology modeling and ran-dom forest classifier. Remote Sens. Environ. 2020, 238, 111017. [Google Scholar] [CrossRef]
  17. Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image seg-mentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, Munich, Germany, 5–9 October 2015; Springer: Cham, Switzerland, 2015; pp. 234–241. [Google Scholar]
  18. Mohajerani, S.; Saeedi, P. Cloud-Net: An end-to-end cloud detection algorithm for Landsat 8 imagery. In Proceedings of the IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 28 July–2 August 2019; pp. 1029–1032. [Google Scholar]
  19. Ye, H.; Liu, S.; Jin, K.; Cheng, H. CT-UNet: An Improved Neural Network Based on U-Net for Building Segmenta-tion in Remote Sensing Images. In Proceedings of the 2020 25th International Conference on Pattern Recognition (ICPR), Milan, Italy, 10–15 January 2021; pp. 166–172. [Google Scholar] [CrossRef]
  20. He, N.; Fang, L.; Plaza, A. Hybrid first and second order attention Unet for building segmentation in re-mote sensing images. Sci. China Inf. Sci. 2020, 63, 1–12. [Google Scholar] [CrossRef]
  21. Jiao, L.; Huo, L.; Hu, C.; Tang, P. Refined UNet: UNet-based refinement network for cloud and shadow precise segmentation. Remote Sens. 2020, 12, 2001. [Google Scholar] [CrossRef]
  22. Hou, Y.; Liu, Z.; Zhang, T.; Li, Y. C-Unet: Complement UNet for remote sensing road extraction. Sensors 2021, 21, 2153. [Google Scholar] [CrossRef] [PubMed]
  23. Ashish, V.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Po-losukhin, I. Attention is all you need. In Advances in Neural Information Processing Systems; Neural Info Process SYS: La Jolla, CA, USA, 2017; Volume 30. [Google Scholar]
  24. Khan, S.; Naseer, M.; Hayat, M.; Zamir, S.W.; Khan, F.S.; Shah, M. Transformers in vision: A survey. ACM Comput. Surv. 2021, 54, 200. [Google Scholar] [CrossRef]
  25. Ze, L.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; Guo, B. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Montreal, QC, Canada, 11–17 October 2021; pp. 10012–10022. [Google Scholar]
  26. Alexey, D.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv 2020, arXiv:2010.11929. [Google Scholar]
  27. Xie, E.; Wang, W.; Yu, Z.; Anandkumar, A.; Alvarez, J.M.; Luo, P. SegFormer: Simple and efficient design for se-mantic segmentation with transformers. Adv. Neural Inf. Processing Syst. 2021, 34, 12077–12090. [Google Scholar]
  28. Winarso, G.; Budhiman, S. The potential application of remote sensing data for coastal study. In Proceedings of the 22nd Asian Conference on Remote Sensing, Singapore, 5–9 November 2001. [Google Scholar]
  29. Feyen, L.; Ciscar, J.C.; Gosling, S.; Ibarreta, D.; Soria, A. (Eds.) Climate Change Impacts and Adaptation in Europe; JRC PESETA IV final report. EUR 30180EN; JRC119178; European Union: Luxembourg, 2020; ISBN 978-92-76-18123-1. [Google Scholar] [CrossRef]
  30. Vos, K.; Harley, M.D.; Splinter, K.D.; Walker, A.; Turner, I.L. Beach slopes from satellite-derived shorelines. Geophys. Res. Lett. 2020, 47, e2020GL088365. [Google Scholar] [CrossRef]
  31. Gao, B.C. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 1996, 58, 257–266. [Google Scholar] [CrossRef]
  32. Xu, H. Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. Int. J. Remote Sens. 2006, 27, 3025–3033. [Google Scholar] [CrossRef]
  33. DeFries, R.S.; Townshend, J.R.G. NDVI-derived land cover classifications at a global scale. Int. J. Remote Sens. 1994, 15, 3567–3586. [Google Scholar] [CrossRef]
  34. Gashaw, T.; Dile, Y.T.; Worqlul, A.W.; Bantider, A.; Zeleke, G.; Bewket, W.; Alamirew, T. Evaluating the effectiveness of best management practices on soil erosion reduction using the SWAT Model: For the case of Gumara watershed, Abbay (Upper Blue Nile) Basin. Environ. Manag. 2021, 68, 240–261. [Google Scholar] [CrossRef] [PubMed]
  35. Phinzi, K.; Ngetar, N.S. The assessment of water-borne erosion at catchment level using GIS-based RUSLE and remote sensing: A review. Int. Soil Water Conserv. Res. 2019, 7, 27–46. [Google Scholar] [CrossRef]
  36. Oldeman, L.R.; van Lynden, G.W.J.; van Engelen, V.W.P. An international methodology for soil degradation assessment and for a soils and terrain digital database (SOTER). In Proceedings of the Paper Presented at the Regional Workshop on Assessment and Monitoring of Land Degradation, Cairo, Egypt, 13–15 May 1995; International Soil Reference and Information Center (ISRIC): Wageningen, The Netherlands, 1995. [Google Scholar]
  37. Lal, R. Water management in various crop production systems related to soil tillage. Soil Tillage Res. 1994, 30, 169–185. [Google Scholar] [CrossRef]
  38. Speth, J.G. Towards an effective and operational International Convention on Desertification. In Proceedings of the International Convention on Desertification, Paris, France, 6–17 June 1994. [Google Scholar]
  39. Pimentel, D. Soil erosion: A food and environmental threat. Environ. Dev. Sustain. 2006, 8, 119–137. [Google Scholar] [CrossRef]
  40. Renard, K.G.; Foster, G.A.; Weesies, D.A.; Mccool, D.K.; Yoder, D.C. Predicting Soil Erosion by Water: A Guide to Conservation Planning with the Revised Universal Soil Loss Equation (RUSLE); USDA: Washington, DC, USA, 1997; ISBN 0160489385. [Google Scholar]
  41. Helz, R.L. Monitoring Ground Deformation from Space; US Department of the Interior, US Geological Survey: Reston, VA, USA, 2005. [Google Scholar]
  42. Barzehkar, M.; Parnell, K.E.; Soomere, T.; Dragovich, D.; Engström, J. Decision support tools, systems and indices for sustainable coastal planning and management: A review. Ocean. Coast. Manag. 2021, 212, 105813. [Google Scholar] [CrossRef]
  43. Rovithis, E.; Makra, K.; Savvaidis, A.; Kirtas, E. The accelerometric network of the Indes-Musa project in the Kalochori area: Configuration, documentation and preliminary data interpretation. Bull. Geol. Soc. Greece 2016, 50, 1100–1119. [Google Scholar] [CrossRef]
  44. Gorelick, N. Google earth engine. In EGU General Assembly Conference Abstracts; American Geophysical Union: Vienna, Austria, 2013; Volume 15, p. 11997. [Google Scholar]
  45. Zuhlke, M.; Fomferra, N.; Brockmann, C.; Peters, M.; Veci, L.; Malik, J.; Regner, P. SNAP (sentinel application platform) and the ESA sentinel 3 toolbox. In Proceedings of the Sentinel-3 for Science Workshop, Venice, Italy, 2–5 June 2015; Volume 734, p. 21. [Google Scholar]
  46. Drusch, M.; Del Bello, U.; Carlier, S.; Colin, O.; Fernandez, V.; Gascon, F.; Bargellini, P. Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote Sens. Environ. 2012, 120, 25–36. [Google Scholar] [CrossRef]
  47. Roy, D.P.; Wulder, M.A.; Loveland, T.R.; Woodcock, C.E.; Allen, R.G.; Anderson, M.C.; Zhu, Z. Landsat-8: Science and product vision for terrestrial global change research. Remote Sens. Environ. 2014, 145, 154–172. [Google Scholar] [CrossRef]
  48. Torres, R.; Snoeij, P.; Geudtner, D.; Bibby, D.; Davidson, M.; Attema, E.; Rostan, F. GMES Sentinel-1 mission. Remote Sens. Environ. 2012, 120, 9–24. [Google Scholar] [CrossRef]
  49. Filipponi, F. Sentinel-1 GRD preprocessing workflow. In Proceedings of the the 3rd International Electronic Conference on Remote Sensing, Online, 22 May–5 June 2019; Volume 18, p. 11. [Google Scholar]
  50. Lin, T.Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 2980–2988. [Google Scholar]
  51. Zha, Y.; Gao, J.; Ni, S. Use of normalized difference built-up index in automatically mapping urban areas from TM imagery. Int. J. Remote Sens. 2003, 24, 583–594. [Google Scholar] [CrossRef]
  52. Terkenli, T.S. Landscape research in Greece: An overview. Belg. Rev. Belg. De Géographie 2004, 2–3, 277–288. [Google Scholar] [CrossRef]
  53. Wischmeier, W.H.; Smith, D.D. Predicting Rainfall Erosion Losses: A Guide to Conservation Planning; Science and Education Administration; U.S. Department of Agriculture No. 537: Washington DC, USA, 1978. [Google Scholar]
  54. Kouli, M.; Soupios, P.; Vallianatos, F. Soil erosion prediction using the Revised Universal Soil Loss Equation (RUSLE) in a GIS framework, Chania, Northwestern Crete, Greece. Environ. Geol. 2009, 57, 483–497. [Google Scholar] [CrossRef]
  55. Chuenchum, P.; Xu, M.; Tang, W. Estimation of soil erosion and sediment yield in the lancang-mekong river using the modified revised universal soil loss equation and GIS techniques. Water 2020, 12, 135. [Google Scholar] [CrossRef]
  56. Zhou, Q.; Yang, S.; Zhao, C.; Cai, M.; Ya, L. A Soil Erosion Assessment of the Upper Mekong River in Yunnan Province, China A Soil Erosion Assessment of the Upper Mekong River in Yunnan Province, China. Mt. Res. Dev. 2014, 34, 36–47. [Google Scholar] [CrossRef]
  57. De Jong, S.M. Applications of Reflective Remote Sensing for Land Degradation Studies in a Mediterranean Environment; Faculteit Ruimtelijke Wetenschappen Universiteit Utrecht: Utrecht, The Netherlands, 1994. [Google Scholar]
  58. Panagos, P.; Borrelli, P.; Meusburger, K.; van der Zanden, E.H.; Poesen, J.; Alewell, C. Modelling the effect of support practices (P-factor) on the reduction of soil erosion by water at European scale. Environ. Sci. Policy 2015, 51, 23–34. [Google Scholar] [CrossRef]
  59. Tian, P.; Zhu, Z.; Yue, Q.; He, Y.; Zhang, Z.; Hao, F.; Guo, W.; Chen, L.; Liu, M. Soil erosion assessment by RUSLE with improved P factor and its validation: Case study on mountainous and hilly areas of Hubei Province, China. Int. Soil Water Conserv. Res. 2021, 9, 433–444. [Google Scholar] [CrossRef]
  60. Huffman, G.J.; Stocker, E.F.; Bolvin, D.T.; Nelkin, E.J.; Tan, J. GPM IMERG Final Precipitation L3 1 Month 0.1 Degree × 0.1 Degree V06; Goddard Earth Sciences Data and Information Services Center (GES DISC): Greenbelt, MD, USA, 2019. [Google Scholar]
  61. EU-DEM. Available online: (accessed on 29 August 2022).
  62. European Commission. The European Soil Database Distribution Version 2.0; European Commission: Brussels, Belgium, 2004. [Google Scholar]
  63. Panagos, P. The European soil database. GEO Connex. 2006, 5, 32–33. [Google Scholar]
  64. Food and Agriculture Organization; International Soil Reference and Information Centre. Jrc: Harmonized World Soil Database (Version 1.2); FAO: Rome, Italy; IIASA: Laxenburg, Austria, 2012. [Google Scholar]
  65. Prandi, P.; Meyssignac, B.; Ablain, M.; Spada, G.; Ribes, A.; Benveniste, J. Local sea level trends, accelerations and uncertainties over 1993–2019. Sci. Data 2021, 8, 1–12. [Google Scholar] [CrossRef]
  66. Rosen, P.A.; Hensley, S.; Joughin, I.R.; Li, F.K.; Madsen, S.N.; Rodriguez, E.; Goldstein, R.M. Synthetic Aperture Radar Interferometry. Proc. IEEE 2000, 88, 333–382. [Google Scholar] [CrossRef]
  67. Zhang, P.; Ke, Y.; Zhang, Z.; Wang, M.; Li, P.; Zhang, S. Urban land use and land cover classification using novel deep learning models based on high spatial resolution satellite imagery. Sensors 2018, 18, 3717. [Google Scholar] [CrossRef] [PubMed]
  68. Rakhlin, A.; Davydow, A.; Nikolenko, S. Land cover classification from satellite imagery with u-net and lovász-softmax loss. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, Salt Lake City, UT, USA, 18–22 June 2018; pp. 262–266. [Google Scholar]
  69. Ulmas, P.; Liiv, I. Segmentation of satellite imagery using u-net models for land cover classification. arXiv 2020, arXiv:2003.02899. [Google Scholar]
  70. Bazi, Y.; Bashmal, L.; Rahhal, M.M.A.; Dayil, R.A.; Ajlan, N.A. Vision transformers for remote sensing image classification. Remote Sens. 2021, 13, 516. [Google Scholar] [CrossRef]
  71. Chen, H.; Qi, Z.; Shi, Z. Remote sensing image change detection with transformers. IEEE Trans. Geosci. Remote Sens. 2021, 60, 1–14. [Google Scholar] [CrossRef]
  72. Ding, L.; Lin, D.; Lin, S.; Zhang, J.; Cui, X.; Wang, Y.; Bruzzone, L. Looking outside the window: Wide-context transformer for the semantic segmentation of high-resolution remote sensing images. IEEE Trans. Geosci. Remote Sens. 2022, 60, 1–13. [Google Scholar] [CrossRef]
  73. De Girolamo, A.M.; Lo Porto, A.; Pappagallo, G.; Tzoraki, O.; Gallart, F. The Hydrological Status Concept: Application at a Temporary River (Candelaro, Italy). River Res. Appl. 2015, 31, 892–903. [Google Scholar] [CrossRef]
  74. Oueslati, O.; De Girolamo, A.M.; Abouabdillah, A.; Kjeldsen, T.R.; Lo Porto, A. Classifying the flow regimes of Mediterranean streams using multivariate analysis. Hydrol. Process. 2015, 29, 4666–4682. [Google Scholar] [CrossRef]
  75. Abdelwahab, O.M.M.; Ricci, G.F.; De Girolamo, A.M.; Gentile, F. Modelling soil erosion in a Mediterranean watershed: Comparison between SWAT and AnnAGNPS models. Environ. Res. 2018, 166, 363–376. [Google Scholar] [CrossRef]
  76. Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735–1780. [Google Scholar] [CrossRef]
  77. Wang, G.; Li, P.; Li, Z.; Ding, D.; Qiao, L.; Xu, J.; Wang, H. Coastal dam inundation assessment for the yellow river delta: Measurements, analysis and scenario. Remote Sens. 2020, 12, 3658. [Google Scholar] [CrossRef]
Figure 1. Study area (red polygon). The yellow polygon shows the location of the study area (main map) in and around of Greece.
Figure 1. Study area (red polygon). The yellow polygon shows the location of the study area (main map) in and around of Greece.
Jmse 10 01322 g001
Figure 2. The proposed platform architecture.
Figure 2. The proposed platform architecture.
Jmse 10 01322 g002
Figure 3. Machine Learning Pipeline.
Figure 3. Machine Learning Pipeline.
Jmse 10 01322 g003
Figure 4. Areas selected for the Training and Testing Sets (best viewed in colour).
Figure 4. Areas selected for the Training and Testing Sets (best viewed in colour).
Jmse 10 01322 g004
Figure 5. Yearly average map composition workflow.
Figure 5. Yearly average map composition workflow.
Jmse 10 01322 g005
Figure 6. Flowchart for the average soil loss estimation using the RUSLE model.
Figure 6. Flowchart for the average soil loss estimation using the RUSLE model.
Jmse 10 01322 g006
Figure 7. Data used in the application of the RUSLE model and the corresponding sources (Precipitation data [60], DEM [61], Land Use [8], Soil Dataset [62,63], Harmonized World Soil Database [64], NDVI index [33]).
Figure 7. Data used in the application of the RUSLE model and the corresponding sources (Precipitation data [60], DEM [61], Land Use [8], Soil Dataset [62,63], Harmonized World Soil Database [64], NDVI index [33]).
Jmse 10 01322 g007
Figure 8. Permanent GNSS station equipment (left) and GNSS receiver antenna (right).
Figure 8. Permanent GNSS station equipment (left) and GNSS receiver antenna (right).
Jmse 10 01322 g008
Figure 9. Installed tide gauge (left) in the commercial port of Thessaloniki (Pier 6, right).
Figure 9. Installed tide gauge (left) in the commercial port of Thessaloniki (Pier 6, right).
Jmse 10 01322 g009
Figure 10. Land Cover Classification with U-Net using the colour coding of Table 1.
Figure 10. Land Cover Classification with U-Net using the colour coding of Table 1.
Jmse 10 01322 g010
Figure 11. Land Cover Classification with modified SegFormer using the colour coding of Table 1.
Figure 11. Land Cover Classification with modified SegFormer using the colour coding of Table 1.
Jmse 10 01322 g011
Figure 12. Ground Truth, WorldCover Layer using the colour coding of Table 1.
Figure 12. Ground Truth, WorldCover Layer using the colour coding of Table 1.
Jmse 10 01322 g012
Figure 13. Generated yearly Average maps (water probability is shown as shades of blue).
Figure 13. Generated yearly Average maps (water probability is shown as shades of blue).
Jmse 10 01322 g013
Figure 14. Average 5-Year flood probability map samples for Evros river delta.
Figure 14. Average 5-Year flood probability map samples for Evros river delta.
Jmse 10 01322 g014
Figure 15. Average seasonal soil loss for the Axios catchment area during winter (left) and summer (right).
Figure 15. Average seasonal soil loss for the Axios catchment area during winter (left) and summer (right).
Jmse 10 01322 g015
Figure 16. Average seasonal soil loss for the Aliakmon catchment area during winter (left) and summer (right).
Figure 16. Average seasonal soil loss for the Aliakmon catchment area during winter (left) and summer (right).
Jmse 10 01322 g016
Figure 17. A sample land subsidence velocity map of the study area for the reference years of 2014 to 2016.
Figure 17. A sample land subsidence velocity map of the study area for the reference years of 2014 to 2016.
Jmse 10 01322 g017
Figure 18. Locations of the selected ground points (1-19) within the study area that were measured in the GNSS survey.
Figure 18. Locations of the selected ground points (1-19) within the study area that were measured in the GNSS survey.
Jmse 10 01322 g018
Figure 19. Real-Time tide gauge measurements.
Figure 19. Real-Time tide gauge measurements.
Jmse 10 01322 g019
Figure 20. Preliminary version of WebGIS Interface: (a) User Menu (b), RGB image composite from Sentinel-2, (c) Urban area change map between years 1990 and 2018 (green: gain, red: loss, white: no change). The layer widget for selecting viewing options for background/foreground layers is also shown.
Figure 20. Preliminary version of WebGIS Interface: (a) User Menu (b), RGB image composite from Sentinel-2, (c) Urban area change map between years 1990 and 2018 (green: gain, red: loss, white: no change). The layer widget for selecting viewing options for background/foreground layers is also shown.
Jmse 10 01322 g020
Table 1. WorldCover class colour-coding.
Table 1. WorldCover class colour-coding.
ColourJmse 10 01322 i001Jmse 10 01322 i002Jmse 10 01322 i003Jmse 10 01322 i004Jmse 10 01322 i005Jmse 10 01322 i006Jmse 10 01322 i007Jmse 10 01322 i008
ClassTreesShrublandGrasslandCroplandBuilt-UpBarren /Sparse VegetationOpen WaterHerbaceous
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Tzepkenlis, A.; Grammalidis, N.; Kontopoulos, C.; Charalampopoulou, V.; Kitsiou, D.; Pataki, Z.; Patera, A.; Nitis, T. An Integrated Monitoring System for Coastal and Riparian Areas Based on Remote Sensing and Machine Learning. J. Mar. Sci. Eng. 2022, 10, 1322.

AMA Style

Tzepkenlis A, Grammalidis N, Kontopoulos C, Charalampopoulou V, Kitsiou D, Pataki Z, Patera A, Nitis T. An Integrated Monitoring System for Coastal and Riparian Areas Based on Remote Sensing and Machine Learning. Journal of Marine Science and Engineering. 2022; 10(9):1322.

Chicago/Turabian Style

Tzepkenlis, Anastasios, Nikos Grammalidis, Christos Kontopoulos, Vasiliki Charalampopoulou, Dimitra Kitsiou, Zoi Pataki, Anastasia Patera, and Theodoros Nitis. 2022. "An Integrated Monitoring System for Coastal and Riparian Areas Based on Remote Sensing and Machine Learning" Journal of Marine Science and Engineering 10, no. 9: 1322.

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop