MAPPING OF WATER QUALITY DYNAMICS DERIVED FROM HYPERSPECTRAL EARTH OBSERVATION DATA FOR THE TAM GIANG LAGOON IN CENTRAL VIETNAM

Introduction

Water is essential for life, and its quality is fundamental to the health and balance of the ecosystems that depend on it. Inland water bodies, in particular, are highly susceptible to changing environmental conditions such as climate change, development pressures, and shifts in land use and cover. These factors pose significant risks to water quality, impacting biodiversity, human health, and overall ecosystem sustainability (Ogashawara et al., 2017). Moreover, water quality has been identified as a critical target for the United Nations Sustainable Development Goals, underscoring its importance in global environmental and public health discourse (Goal 6 | Department of Economic and Social Affairs, 2024n.d.).

Satellite remote sensing offers a promising technology for monitoring water quality, particularly in inland water bodies. While this approach is well-established in ocean monitoring, its application to inland waters presents unique challenges and opportunities. The current generation of satellites dedicated to water monitoring, such as the recently launched PACE on March 2024, are primarily focused on oceanographic applications due to their larger pixel sizes, which limit their effectiveness for inland water monitoring (Dierssen et al., 2023). As a result, many studies have relied on multispectral satellites originally designed for terrestrial applications. However, the limited band resolution of these satellites often hinders precise water quality retrieval (Giardino et al., 2019).

To address these challenges, researchers have increasingly turned to hyperspectral satellite imagery, which provides detailed spectral information across a wide range of wavelengths with narrow spectral bands. This capability enables the differentiation of various water quality parameters and algae types that multispectral sensors may miss (Gholizadeh et al., 2016a). Hyperspectral sensors such as EnMAP, and DESIS have been employed in this study to retrieve detailed water quality information for the Tam Giang Lagoon.

The methods implemented for water quality retrieval in this thesis focus on physics-based models, which provide more coherent and consistent results over time compared to traditional band ratio or index methods. Specifically, the Water Colour Simulator (WASI) model is used to interpret hyperspectral data, allowing for accurate assessment of key water quality indicators such as chlorophyll-a, total suspended solids (TSS), and colored dissolved organic matter (CDOM). The WASI model leverages spectral inversion techniques to provide robust estimates of these parameters, overcoming the limitations of empirical models (Gege, 2014).

In addition to data retrieval, the visualization of water quality dynamics plays a crucial role in communicating findings and informing management decisions. This thesis explores various geovisualization techniques to represent the spatial and temporal variations in water quality effectively. Advanced cartographic visualization methods are employed to create intuitive and interactive maps, enhancing the interpretation of complex datasets (MacEachren & Kraak, 2001). By integrating geovisualization tools, the research aims to present water quality data in a way that is accessible and actionable for local authorities and stakeholders in the Hue region.

My personal interest in this topic stems from growing concerns about water resources and contamination in my local environment. I have observed firsthand the difficulties in tackling water pollution issues, flash floods, and droughts. I am motivated by the potential for this research to have a direct impact on the Hué region and the inhabitants that make a living from the Tam Giang lagoon. The results of this thesis could provide valuable insights for authorities in the Hue region and potentially offer geovisualization tools to inform and engage the local population.

Objectives

The primary objective of this master thesis is to analyze the temporal evolution of water quality in the Tam Giang, Lagoon using time series hyperspectral EO data and to design advanced geovisualization techniques to better visualize the results.

There are two aspects addressed by the thesis assignment: the analysis of water quality parameters over time, and the visualization and cartographic representation of the derived results. Therefore, there are two research parts in the thesis: the thematic analytical part presented as Research 1 and the geovisualization part presented as Research 2.

• The theoretical aim of the thesis is to make a comprehensive literature review focusing on the topics of water quality indicators, hyperspectral EO data, and methods of cartographic visualization of obtained results.
• In the Research 1 part there are explored hyperspectral (EnMap, DESIS, PRISMA) EO sensor data to assess critical water quality indicators in the lagoon. The sub-goal of the thesis is to analyze and subsequently understand some of the dynamic behavior of the ecosystem. A temporal analysis is therefore used to discern patterns and dynamics in water quality parameters. The study has to assess variations in water quality indicators over time, shedding light on the lagoon's environmental changes and temporal dynamics.
• In the Research 2 part there is an analysis of possible geovisualization of obtained results. Innovative geospatial mapping visualization techniques would effectively illustrate outcomes from the thematic part (Research 1).
• The expected outcomes of the theses have to present the analysis and advanced geovisualisations of the Lagoon water quality dynamics.

Tam Giang Lagoon

The Tam Giang Coastal Lagoon is the largest lagoon in Southeast Asia Figure 2 covering an area of 216 km2. It has a mean depth of 2 m and a maximum depth of 5–15 m in the inlets region caused by the water fluxes(Nghiem et al., 2007). The Lagoon has a sandbar of 66 km long and up to 20 m high; this acts as a barrier between the lagoon and the Sea. The barrier is crossed by inlets which change through time. Currently, there are two of them, called Thuan An inlet in the north and Tu Hien inlet in the south. The opening of these channels and their dynamics are closely related to the rainy season and flood season, which have led to major changes recorded for the lagoon dynamics (Tran Thanh, 2011). The lagoon system receives water from the Huong River basin and is the main contributor to the lagoon (Tran Thanh, 2011). The population along the lagoon is about 300,000 inhabitants, belonging to 5 districts and towns: Phong Dien, Quang Dien, Huong Tra, Phu Vang and Phu Loc (Trang et al., 2013).

During the rainy season, from mid-October through February, the lagoon experiences the main influx of water from the previously mentioned river. Moreover, another water addition in the lagoon is in the Typhoon season from early September through November, these intense precipitation episodes can occasionally lead to flooding. During these two periods more than 70% of the flow volume per year is attributed to the lagoon (Nghiem et al., 2007). This intense precipitations also causes higher turbidity in the lagoon and the decrease in the salinity (Luong et al., 2020). Most of the additions of water are discharged in the southern half of the lagoon.

Methodology

The workflow developed in this thesis Figure 3 starts with the data selection and preprocessing. The EnMAP and DESIS data are selected without fine clouds, Georeferenced in ArcGIS Pro, and the images are converted to specific format file .hdr ENVI format customized in Notepad++. Following the guidance of the WASI manual and taking into account the model ambiguities, the satellite hyperspectral images from the sensors EnMAP and DESIS with the inversion model for deep waters were processed, estimating three water quality parameters TSS, CDOM, and green algae content. After the water quality of the satellite data was gathered, it was compared statistically to the in-situ data; the data that was common was the TSS. From the satellite images the information for the analysis was extracted with ArcGIS Pro and then the statistical analysis was conducted with Python libraries and executed in Google Colab. Also, it was analyzed the relationship of the in-situ data, specifically the TSS vs the rest of the parameters, using Python libraries and Google Colab. With all the data and information, it was visualized the data with static maps utilizing ArcGIS Pro and Adobe Illustrator for a customized composition. Then, it was created an animation for better comprehensiveness of the time series using the static layers created previously and incorporating them in Adobe After Effects using the plugin ArcGIS Maps for Adobe Creative Cloud. Every following section explains more in detail the methodology.

Results

The Following chapter gathers all the final outputs of the thesis. In the first part, there are the results of the analysis part with the outputs of the water quality parameters from the model and the statistics for the satellite validation. In the second part, geovisualization results are presented in the static visualization and animation.

Results of the watr quality analysis

The following water quality parameter results are performed with the inversion modeling of WASI6 2D and the AI plugin. The model used for the processing is the deep-water model. Even though the lagoon had a mean depth of 2 meters, the deep-water model was giving more robust results due to the ununiformed of the lagoon base and the reflectance that this was providing.

The following images show a sample of the processing results (for the entire results, go to attachments). There are three types of data. 1) The resulting image with the water quality parameter and two different graphs indicates 2) the estimation of the model's calibration before training and 3) the accuracy of the model's results once it is trained with water quality parameters. These three types of results are obtained for all the images processed.

Figure 45, Figure 48, Figure 50 show the fitting parameters for the calibration of the spectrum trend of the pixel selected in color blue and the model spectrum with the parameters in red. The verification of the accuracy of the training and the modeling of the parameters is in red and the regression expected to fit is in blue Figure 46, Figure 49, Figure 51. If the two trends are close to each other, it means that the parameter can be expected as a good result, as the figures mentioned above show, the parameter of algae is the less accurate.

As you can see in the fit parameters, these inversions are done using a model for deep waters. This model performed better for the parameter characterization Table 1; the only parameter that underperformed was the algae content. The model used does not take into account the bottom reflectance of the lagoon. As you can see in Table 2 below, the R square errors that are lower than 0.40 have not been taken into account as good results and are highlighted in red. The table represents the accuracy of the results for every image and parameter. The visualization of every result can be seen in the attachments.

Parameters obtained in the results from WASI manual:

Water quality parameter results for EnMAP

The processing of EnMAP data was performed with two different models. The first model tried was the shallow water inversion with WASI software and the second model tried was the deep-water model with WASI-AI. The method that had better results was the deep-water model that is explained in the 4.4 chapter in the justification of methods. The results obtained and analyzed were from the deep-water model. In the thesis document attachments you can find the results.

The first attempt of image processing was processed using the shallow waters model due to the shallow depth of the lagoon which has an average of 2m depth. Also, during the first inversion the fitting parameters of water depth temperature and diatom, the zB (depth), the temperature parameter did not fit the FA(2) and the FA(2). The results were not as accurate as expected and the second attempt proceeded. Peter suggested to process the data with the deep-water inversion because the color of the water and the benthonic organisms that are in the bottom of the lagoon were difficult to quantify.

On the following lines you can find an example of the two attempts, for all the results consult the attachments.

Results of first testing EnMAP model of shallow waters:

Results of analysis EnMAP Model of deep water:

Water quality parameter results for DESIS

The DESIS data has been processed using the deep-water model. Below, you can find an example of 3 of May 2023 the same date as the example of EnMAP data above. For the rest of the results, check the attachments.

Results of analysis pixel statistics of every image

The three tables on the document thesis shows the results of all the satellite images giving the basic statistics, of the green algae(b5), TSS(b6) and CDOM(b7). Statistics: maximum, minimum, mode, mean, standard deviation count of pixels, supposed number of outliers, sum, range of data, and the coefficient of variation. All of the graphs of the pixels can be found in the attachments.

Results of the statistical analysis TSS validation

The First results are from the comparative analysis of the in-situ data to the satellite data and the trends of this data. The second results are the ones showing the relationship between the TSS and the rest of the parameters of the in-situ data.

Results of the comparative in-situ vs satellite TSS

concentration

Descriptive Statistics V2:

In-situ Data:

• The mean TSS is 12.20 mg/L, with a standard deviation of 5.59 mg/L.
• The TSS values range from 5.6 mg/L to 27.20 mg/L.
• The interquartile range (IQR) shows that 50% of the TSS values fall between 7.10 mg/L and 14.80 mg/L.

Satellite Data:

• The mean TSS is 7.15 mg/L, with a much higher standard deviation of 6.64mg/L.
• The TSS values range from 0.83 mg/L to 35.98 mg/L.
• The IQR shows that 50% of the TSS values fall between 4.45 mg/L and 7.25 mg/L, indicating a wide range of variability.

Pearson Correlation Coefficient: 0.844

• There is a low negative correlation between the in-situ and satellite TSS measurements.
• The p-value is 0.854, indicating that the correlation is not statistically significant.
• The IQR shows that 50% of the TSS values fall between 4.45 mg/L and 7.25 mg/L, indicating a wide range of variability.

Regression Analysis

• The slope of the regression line is -0.038, and the intercept is 7.62.
• The R-squared value is 0.001, meaning that approximately 0.10% of the variability in satellite TSS can be explained by the in-situ TSS measurements.
• This suggests a very weak relationship between the two measurements, with most of the variability unaccounted for by the regression model.

Paired T-test

T-statistic: 3.39

• The t-test yields a t-statistic indicating a significant difference between the mean TSS values from in-situ and satellite measurements.
• The p-value is 0.0018, indicating that this difference is statistically significant.

Wilcoxon Signed-Rank Test

Wilcoxon Signed-Rank Test Statistic: 67.0

• The Wilcoxon signed-rank test yields a very low p-value (1.19e-05), indicating a statistically significant difference between the paired samples.
• This non-parametric test is used to compare paired samples and is robust against non-normality. The significant result suggests that there are differences in the paired data that the t-test might not detect due to its assumptions.

Root Mean Square Error (RMSE)

RMSE: 10.05

• The RMSE of 10.05 indicates the average magnitude of error between the satellite and in-situ TSS measurements.
• This suggests substantial differences between the measurements.

Bland-Altman Plot Analysis

The Bland-Altman plot analysis can be visualized on Figure 56 and it shows the following data:

Mean Difference: 5.05

• This result indicates that, on average, the in-situ TSS measurements are higher than the satellite TSS measurements.
• The small mean difference points out that there is no substantial systematic bias between the two measurement methods. The mean difference points out a potential systematic bias where satellite measurements tend to underestimate TSS.

Standard Deviation of Differences: 8.69

• The high standard deviation of the difference and the wide limits of agreement (mean difference ± 1.96 * std_diff) indicate a considerable variability between the individual in-situ and satellite TSS measurements, with potential outliers.
• This reinforces the finding from the RMSE, showing substantial discrepancies between the two methods on a point-by-point basis.

Plots of all the sample points from the in-situ data, containing the in-situ data of the point and also the satellite data of the point though time. Find in the following lines there are 3 examples of the plots Figure 57, Figure 58, and Figure 59, for seeing the rest look in the attachments.

Results of how many times the values have exceeded the 50mg/l from the in-situ data and satellite data Figure 60.

Results comparison of all the parameter vs TSS in

in-situ data

T-test has been realized by comparing the TSS to the rest of the parameters.

Interpreting Small P-values: A very small p-value (typically less than 0.05) indicates strong evidence against the null hypothesis, suggesting that there is a statistically significant difference between the compared groups.

The Shapiro–Wilk test has been used to choose the correlation method to use. In that case, the p-value indicates that the data is not normally distributed; the value is very small.

Shapiro-Wilk Test for TSS: p-value=2.9130588875693885e-36

The proper correlation coefficient to use is the Spearman Correlation Coefficient. The relationship between the total values of the TSS and the particular water parameters has been calculated and shown.

Values between -0.3 and 0.4 indicate that the correlations between TSS and the other parameters are weak to moderate. This means that the relationship is either weakly linear or potentially non-linear.

• -0.3 to -0.1: Weak negative correlation. As one parameter increases, TSS slightly decreases.
• -0.1 to 0.1: Very weak or no correlation. There is little to no linear relationship between TSS and the parameter.
• 0.1 to 0.3: Weak positive correlation. As one parameter increases, TSS slightly increases.
• 0.3 to 0.4: Moderate positive correlation. As one parameter increases, TSS moderately increases.

The Spearman correlation shows a moderate positive correlation for CDO and Fe and a moderate negative correlation for DO.

Results of the geovisualization

The results of the geovisualization are a series of static maps and a map animation.

Static maps:

The 4 map posters Figure 62, Figure 63, Figure 64, Figure 65, utilize the method with nonclassified color scales. Visualizing the temporal evolution of the water quality through time. Also, it uses a more advanced method that transforms the raster into a regular grid and classifies the color scale into different segments. Also, it is done some statistical measurements and analysis to summarize the period into a synthetic map showing the differences between the minimum and the maximum values of the period.

Map animation:

For the animation, it is selected to visualize the TSS parameter through time Figure 66, Figure 67 because it is the only parameter that it has registered for the satellite data, and it also contains credible in-situ results. Also, as summary is visualized Figure 68, the synthetic map is also done for the map posters, showing the differences between the minimum and the maximum values of the period. In the animation it is also incorporated the data of the flooding mentioned in the visualization data chapter. This layer has been visualized in a simple color scale for the TSS of in-situ and satellite data, and the flooding represents a uniform color for the delimitation of the event.

Conclusions

Water quality in inland bodies is a critical environmental indicator and resource that provides essential ecosystem services, including sustaining biodiversity, supporting agriculture, and enabling recreational activities. The increasing levels of pollutants and the impact of human activities on water bodies have prompted significant research into water quality dynamics. Despite advancements in environmental monitoring, Tam Giang Lagoon remains under-monitored for its size, accessibility challenges, and limited technological infrastructure. The Tam Giang Lagoon is a vital ecosystem, representing the largest coastal lagoon in Southeast Asia, yet it faces threats from pollution and sedimentation that require continuous monitoring and management. In this context, Earth Observation (EO) technologies, combined with hyperspectral data analysis, offer a promising approach to enhance our understanding and management of such complex aquatic ecosystems.

Motivated by this need, this thesis focused on developing a comprehensive methodology for assessing water quality parameters in the Tam Giang Lagoon using hyperspectral remote sensing images of EnMAP, and DESIS from 2021 to 2023 and inversion models. The objective was to retrieve and analyze key water quality indicators, including TSS, CDOM, and green algae concentrations, by experimenting with the capabilities of WASI-AI software based on neural network training and the WASI-2D module. The thesis also aimed to explore the geovisualization techniques to effectively communicate these complex datasets to both expert and non-expert audiences.

The water quality analysis revealed significant spatial patterns and temporal trends in TSS and CDOM, indicating areas on the central-north of the lagoon and the borders of the lagoon, respectively, of concern and potential management focus within the lagoon for higher concentrations. Despite the challenge faced in retrieving accurate data for green algae concentrations due to the model, the study successfully highlighted the importance of atmospheric corrections and model calibration for enhancing the accuracy of remote sensing data in aquatic environments. The statistical analysis underscored the complexity of correlating satellite data with in-situ measurements, emphasizing the need for data validation of the same data for a robust water quality assessment.

In terms of geovisualization, the development of static maps and animations provided a multifaceted approach to disseminating water quality information. The hexagonal tessellation method facilitated a clear representation of water quality parameters, while animations offered an engaged way to visualize changes over time. The visualization outputs were tailored for specific target groups, applying techniques of user-centric design with detailed maps for experts to aid in decision-making and simplified and guided animations for the general public to raise awareness of the lagoon's dynamics.

In conclusion, this thesis represents a significant step in utilizing Hyperspectral EO and geovisualization to enhance water quality monitoring in the Tam Giang Lagoon. It demonstrates the potential of combining EO technologies with in-situ data analysis and visualization techniques to provide actionable insights for aquatic ecosystems. As the methodologies and tools developed in this study are open and adaptable, they can be extended and refined for future research and applied in other regions. It is a fact that this work will contribute to increased awareness and understanding of water quality issues, prompting more informed management and conservation efforts in the Vietnamese entities that it will be shared with. The next steps involve refining the inversion models, validating the accuracy of retrieved data, and expanding the visualization capabilities to incorporate more multivariate maps, real-time data feeds, and interactive web-based platforms if it is agreed with the data providers.

Elements

i = 0;

while (!deck.isInOrder()) {
    print 'Iteration ' + i;
    deck.shuffle();
    i++;
}

print 'It took ' + i + ' iterations to sort the deck.';