KEYWORDS: Moderate Resolution Imaging Spectroradiometer, Sentinel-1, surface water monitoring, Google Earth Engine... The goal of this study is to assess the temporal changes of the wate
Study area
Tonlé Sap lake is a lake in northwest Cambodia, it lies in the Mekong River system Tonlé Sap lake has coordinates between 13.26 ◦ N and 12.52 ◦ N and between longitude 103.69 ◦ E and 104.46 ◦ E Tonlé Sap lake is the largest freshwater lake in Southeast Asia (about 2700 km 2 ) It is one of the most productive inland fisheries in the world Apart from rainfall, the lake receives water from the Mekong River through the Tonlé Sap River This supplies nutrients for the lake and thus the surrounding area is rich in alluvial soil, which is favorable for agricultural development
Figure 1: The map of Tonlé Sap Lake
The area and volume of the lake vary significantly throughout a year The lake is connected to the Mekong River through the Tonlé Sap River at Phnom
The high water levels of Mekong River during rainy season would increase the flow of Tonlé Sap river into the lake therefore increase the lake from 2700 km 2 to 9000-16000 km 2 in area and 1 km 3 to 80 km 3 in volume.[7][10]
The direction of the flow is reversed during the dry season The Mekong River receives water from the lake, which elevates the water levels in the delta for about 5 to 6 months during the dry season.[8]
Annually, Tonlé Sap lake and the inland water systems contributes around 500000 tons of fish [3] It is suggested that overfishing and low precipitation had led to the decrease in fish production [2] Sand mining in the Lower Mekong River is also a cause of the shrinking of Tonlé Sap lake [9].
Data collection
Google Earth Engine (GEE) is an open platform for image processing
GEE provides many image products of various satellites, it is a coding space that allows researchers to process images and download them if needed
MODIS (Moderate Resolution Imaging Spectroradiometer) is a satellite sensor used for earth and climate measurements It is on board the Terra satellite, launched by NASA in 1999
The MODIS product used for this project is MOD09A1.006 (Terra
After adjusting for atmospheric factors such as gases, aerosols, and Rayleigh scattering, the MOD09A1 V6 output gives the estimated surface spectral reflectance of MODIS band 1 to band 7 at 500m resolution Based on factors including good observation coverage, low view angle, lack of clouds or cloud
8 shadow, and aerosol loading, a value is chosen from all of the acquisitions made up of the 8-day composite for each pixel
Band Spectral range(nm) Description
Table 1: Spectral bands of MODIS
The Sentinel-1 images from S1A products that Sentinel-1A collected between 2015 and 2020 were used in this study Sentinel-1 Ground Range Detected (GRD) data that covered the Tonlé Sap Lake in IW mode and VH polarization was downloaded from Copernicus Open Access Hub
The impact of speckle is lessened by using Level-1 GRD, which provides data with detected amplitude and multiple looks Additionally, the information was in Interferometric Wide High- and medium-resolution Swath GRD mode with multi- looking done on every burst separately Then, each polarization channel received a single, contiguous ground range detected image after merging all of the burst from each sub-swath.[16]
Studies have shown that VH polarization gives out high accuracy when used to extract water [15] In this study, only VH polarization is selected for analysis
The analysis results of MODIS images is going to be compared to the temporal water level of the lake collected from Hydroweb and the temporal estimated water discharge measured at two hydrological stations Tan Chau
Hydroweb, a database developed by LEGOS (a joint research unit sponsored by CNES, CNRS, IRD, and UT3), offers access to water level variations in lakes, reservoirs, and rivers Leveraging satellite altimetry, Hydroweb provides free data from various satellites The database encompasses thousands of virtual stations monitored in real-time or delayed mode, providing users with valuable information on water levels globally.
The water level of Tonlé Sap lake is in operational mode, its time series span from 1995 to 2022 For the purpose of this study, the water level of Tonlé Sap lake in the period 2000-2021 is selected for assessment
The discharge measurements at Tan Chau and Chau Doc stations were provided by my supervisor These measurements would be compared to the volume variations calculated from the analysis results.
Methodology
The workflow is demonstrated as below
2.3.1Surface Water Extent Mapping with MODIS Imagery (2000- 2021)
First, on Google Earth Engine platform, import the image product MOD09A1.006, because the product covers globally so it is necessary to import a shapefile of the study area to create a subset later While choosing a time series for the analysis, it is recommended to assess the image collection year by year instead of assessing the two decades as a whole to avoid GEE taking too much time to process
Thirdly, write the functions to calculate MNDWI (Modified Normalized Difference Water Index) and extract the bands for further analysis One notable feature of GEE is that not only the images but also the image collections can be processed in functions
The MNDWI formula is defined as: ýþ �㔷þā = ÿ �㕟�㕒�㕒�㕛 2 þþāý ÿ�㕟�㕒�㕒�㕛 + þþāý
The green band and short-wave infrared band are respectively band 4 and band 6 in MODIS so apply the MNDWI function on these 2 bands accordingly
After layers of the MNDWI map are created, create a subset of the study area with the aforementioned shapefile
Now, each image in the selected image collection would be called out in a for loop to be attached with selected bands, Band 3 (blue band) and MNDWI (can be saved as a band after calculation)
Finally, each image would be exported to Drive in GeoTIFF format, with its date time saved in its filename
The images can be downloaded directly from Google Drive
From now, the analysis process continues in Matlab First, it is necessary to
11 check the dimensions of images Though the process in GEE is the same for all images, but it’s required that they have same dimensions as well to avoid complications later The image’s dimension and other parameters can be preview on ENVI Classic
The images can be opened in Matlab with the geotiffread function To process images consecutively, put all functions in a for loop
Images are represented as 249x139 matrices with each pixel having associated Band 3 (Blue) and MNDWI values The MNDWI value determines the binary value (1 or 0) of each pixel in a corresponding binary image, resulting in the same dimensions as the original image.
Primarily set 0 as the MNDWI threshold value, if the MNDWI value one pixel is greater than 0 then it’s value in binary image is 1, meaning the pixel is classified as water and, vice versa, the pixel would be classified as water
Cloud pixels arise when cloud cover interferes with reflectance from the ground, and their reflectance often exceeds ground reflectance in the VNIR spectral region To identify cloud pixels, researchers have proposed using a threshold value for the blue band.
The spatial and temporal variation in cloud cover can greatly affect regional and localized weather processes There are totally 773 images that need to be corrected in varying degrees
Due to the comparatively high reflection of clouds in this band and the low reflectance of the majority of the earth’s surface, the blue band difference was
12 chosen above other bands The threshold value used in this study is 2000, this means that if a pixel’s B3 value is greater than 2000 then it would be classified as a cloud pixel, and its MNDWI value is miscalculated and would be changed to NaN format to be re-calculated later
• Re-calculating cloud pixels by interpolation
There are totally 1005 images for the whole 2000-2021 period, each images would have 34611 pixels
Firstly, create a zero matrix A with dimension 1005x34611 Next, use a for loop to store all MNDWI values (including NaN values) in this matrix If all images now have changed from 249x139 dimension to be one- dimensional so one row of this matrix would contain all MNDWI values of 1 image
So the transpose matrix A’ would have 1005 rows and 34611 columns, each row now would be a time series of a specific pixel
Figure 3: Image with a small portion of cloud pixels (colored green)
Figure 4: Image with a large portion of cloud pixels, covering mostly of study area (colored green)
13 Applying linear interpolation to each row would replace NaN values (cloud pixels) with acceptable MNDWI values
Linear interpolation formula: þ = þ 1 + (ý 2 ý 1 )þ22 þ1 ý 2 2 ý 1 Below is an example of how interpolation replaces unknown values in an array
Figure 5: Fixing NaN values with interpolation
• Recreating binary images and comparing to reference data
The first step is simply re-running the previous codes with improved MNDWI values The result binary images are expected to show very clear and consistent shape of the lake
The second step is to compute the water surface area of images by multiplying the sum of pixels with 1 km 2 Compare this temporal change of surface water area with the time series of water levels downloaded from Hydroweb
Similarly, compute the water volume variations of the lake by this equation:
After that, compare these variations with the sum of water discharge measurements at two hydrological stations Tan Chau and Chau Doc
The reference datasets were saved in csv or txt format that can be accessed and read in Matlab
2.3.2Surface Water Extent Mapping with SAR Sentinel-1 Imagery (2015-2021)
For SEO optimization, the key sentence in the paragraph should be:**To analyze the Sentinel-1 imagery within the study area, the platform allows for subsetting by importing the shapefile and Sentinel-1 SAR GRD collection Additionally, a larger square-shaped Geometry boundary is imported to cover the Tonlé Sap lake.
Set the research period starting from 2015/01/01 to 2021/12/31
The Speckle filtering function is written to help reduce noise from the interference between many elemetary scatters thus improving image quality.[11]
Otsu threshold function Named after Nobuyuki Otsu, Otsu’s method was developed to automatically find image threshold, in other word, it is an algorithm that returns a threshold value that separates pixels into 2 different classes, foreground and background.[5]
In this project, Otsu’s method would help to create binary images, classifying water and land pixels
With Otsu threshold function, each image would return a different threshold value The histogram function to draw the histogram of each image All processes could be iterated in a loop for all available images in the selected collection So for each image in the collection, there would be a histogram, a threshold value, a binary water mask, and later in Matlab, the water surface area could be computed And finally, a time series of surface water extent to compare with the results of MODIS.
Surface Water Extent Mapping with MODIS Imagery (2000- 2021)
After the first step of processing in Matlab, 1005 images downloaded from
15 GEE would result in 1005 pairs of images, including 1 MNDWI mask and 1 water binary mask
Before cloud filtering, the cloud pixels could greatly affect the two masks , leading to a false sum of water pixels (with value 1 in binary images), leading to a false calculation of surface water area
(a) MNDWI mask, not greatly affected by cloud pixels (b) Binary water mask, water body has a clear shape
(c) MNDWI mask,affected greatly by cloud pixels (d) Binary water mask has many missing pixels
Figure 6:Before cloud filtering (green pixels represents water, blue pixels are non-water)
Variations in water pixel time series result from missing pixels falsely classified as non-water due to erroneous MNDWI values This misclassification arises from cloud cover obscuring these pixels, leading to fluctuations in the graph.
Figure 7: Water pixel count before cloud filtering
Figure 8: Water surface area after cloud filtering
Cloud filtering greatly changed the dataset, the final surface area time series has a clear fluctuating seasonal change, which showed the Mekong’s annual flood pulse
3.1.2Comparing to Hydroweb water level
The strong correlation between the datasets (R = 0.873) suggests a notable agreement between them Both datasets highlight significant increases in August 2011, October 2013, and October 2001 and 2002 Notably, these periods correspond to years with devastating floods in Cambodia, indicating a potential relationship between the data and flood events.
Figure 9: MODIS surface water area (km2) vs Hydroweb water level (m)
These two datasets also have a good correlation with R = 0.861
Apparently the volume variations calculated from image analysis have clear fluctuations compared with the monthly water discharge variations This shows that the image processing, the calculations have correctly reflected the real measurements
Figure 10: MODIS water volume variations (m3) vs discharge measurements(m)
Surface Water Extent Mapping with SAR Sentinel-1 Imagery (2015- 2021)
Although I have successfully coded the Otsu function to give out the expected threshold value, a n d p l o t t e d a c o n s i d e r a b l e n u m b e r o f b i n a r y i m a g e s , t h e r e w e r e m a n y o b s t a c l e s t h a t h i n d e r e d f i n a l a n a l y s i s s u c c e s s First of all, the process on Google Earth Engine is very time-consuming as the image collection for the 2015-2021 is a significantly large dataset, and the codes for the Sentinel-1 analysis is significantly more complicated
Secondly, Tonle Sap lake is a huge study area, and Sentinel-1 does not cover the whole study area in images taken on a same day This leads to a bigger collection of smaller images, leads to many difficulties with sorting out images or merging the smaller images to make a better image that covers the study area sufficiently The total number of images, including those that needs to be sorted out, can get to 500 for the entire
19 period of 2015-2021, this problem demands a better understanding of the dataset to devise a better way for selecting images
Though I have not figured out a way to have a better image collection, the remaining analysis process promised good results
For each Sentinel-1 image acquired from Google Earth Engine, a histogram and a threshold value are employed to differentiate water from non-water pixels The threshold value ranges primarily between T = -27 and T = -20.9, serving as a crucial parameter in the classification process.
Figure 11: Histogram of backscatter VH of Sentinel-1 from Google Earth Engine on 15 May 2015
Figure 12:Binary water mask of Tonle Sap Lake (01 July 2020)
Utilizing optimized threshold values and an enhanced image dataset, the binary masks representing water surface areas can be systematically generated By tracking changes in these masks over time, a temporal analysis of water surface area variations can be conducted, providing valuable insights into hydrological dynamics.
In summary, in this study, I work on the temporal change of the water extent and volume of Tonlé Sap lake
The cloud filtering method effectively enhances the evaluation of temporal water body changes Notably, the MODIS image collection strongly correlates with water level data from Hydroweb (acquired through satellite altimetry) and discharge measurements at hydrological stations This underscores the reliability of the cloud filtering approach in capturing the dynamics of water bodies over time.
Same methods could be applied to the monitoring of other lakes in Vietnam Both methods were executed in Google Earth Engine, which was a very helpful tool that provides free and intensive data analysis The only factor could gravely affect the working process on Google Earth Engine is its connection to server, as it frequently stops running in case of heavy workload
There are also many other different cloud filtering methods that were suggested: for example in [14], the author suggested using the green band instead of blue band for its sensitivity to atmospheric factors like aerosol or smoke This is an interesting idea I would like to investigate and validate for this particular project on Tonle Sap lake
For the future, I would like to finish the Sentinel-1 image analysis on Tonle Sap lake’s water extent as the functions and assessment are finished, leaving the only one problem behind is selecting a better image collection or devising a way to merge the