Floods and landslides are the first and fourth most frequent disasters around the world (Petley, 2012). There are several examples of downstream flooding caused by massive mudslides where rapid mapping is an indispensable tool for supporting disaster management activities by civil protection authorities.
Since July 2014, the Copernicus programme of the European Union has been providing free-of-charge access to Sentinel-1 radar data coveirng the entire world. This allows for the exploration of new applications to strengthen hazard monitoring and disaster mitigation activities.
This UN-SPIDER Recommended Practice emphasizes the use of SAR data during and after a disaster crisis, since optimum atmospheric conditions for optical satellite images are not always available. In the example provided here, it is used to map the mudflow following the dam collapse that occured on 25 January 2019 at Brumadinho, Brazil.
- For applying this methodology, it is necessary to use SAR images in Level-0 (equivalent to RAW product type) or Level-1 (equivalent to SLC product type). According to Big SAR Data available in the Copernicus service platform, the following Sentinel-1 acquisition modes can be used: Stripmap (SM), Interferometric Wide swath (IW) and Extra-Wide swath (EW).
- For the ortho-rectification process, a Digital Elevation Model is necessary, which is automatically downloaded by SNAP software.
For this recommended practice, SNAP is necessary. The software is freely available online for Windows 32 and 64-bit versions and Mac OS X; the current version is 6.0.0.
Similarly, we used QGIS software available for free for Windows 32 and 64-bit versions and Mac OS X; the current version is 3.6.3
Basic understanding about SAR theory and SAR image processing is required.
The minimum computing capacity specifications recommended are the following:
- 300GB of free disk space
- 8GB of RAM
- Dual core processor (Intel i7)
Strengths and Limitations:
This method is recommended when the flooded areas cannot be detected by the traditional techniques due to ground moisture conditions or the large volume of sediments involved during a flood. The strength of this method lies in the digital treatment techniques, which allow the most information about backscattering mechanisms linked to the surficial roughness of the massive mudslide to be extracted. Thus, it is possible to detect spatial patterns of the disaster undetectable by an optical image.
This recommended practice is a useful tool for rapid mapping to support the emergency response during a disaster crisis by civil protection authorities.
This method can be applied only in massive mudslides and associated major flooding. Similarly, to apply this recommended practice only SAR images with two polarizations minimum can be used.
The aim is to identify the area affected by mudslide and associated floods using the backscattering information from "magnitude" SAR data. This step-by-step procedure applies a ratio-change detection technique and principal component analysis using SAR images before and after the dam collapsed.
Disaster type: Landslide and Flood
Disaster Cycle Phase: Recovery & Reconstruction and Relief & Response
Minas Gerais located in Brumadinho, Brazil
This Recommended Practice was developed by a visiting scientist from the Autonomous University of Mexico State (UAEM) who benefitted from the support provided by the National Council of Science and Technology (CONACYT-Mexico). This method was applied for a massive mudflow that occurred 25 January 2019 in Brumadinho, Brazil, where a dam collapsed in the Minas Gerais mining complex triggering a mudslide and causing the death of 232 people. A large volume of sludge flooded at least 12 km downstream, burying mining infrastructure and nearby workers' houses.
This UN-SPIDER Recommended Practice can be applied to all SAR images with dual polarization (HH, VV, HV, and VH) or full polarization (HH, VV, HV, and VH). Single polarization is not recommended to use in this practice since the digital treatment technique needs two polarizations as a minimum requirement.
The methodology can be applied over massive mudslides and associated floods, also when large areas of sludge-water are involved. It is essential to mention that the proposed method is focused on extracting the most significant amount of backscattering information linked to the surficial roughness of the mudslide using the magnitude information of SAR image. To achieve this task, we use change detection techniques (Log- Ratio) and principal component analysis (PCA).
The applicability of PCA to detect massive mudslides (and associated floods) lies in applying digital treatment using the magnitude values of SAR images acquired on different dates. PCA implies a data regrouping, where towards the first "out band components" the variance is maximized, which means that the more significant pixel information is keeping; while towards the last "out band component," all noise or redundant information is separated. This method is highly recommended when the flooded areas cannot be detected by the traditional change detection analysis due to soil moisture conditions, differences between water bodies or by the large volume of sediments transported during a flood.
For this Recommended Practice, we used Sentinel-1 data freely available on the Copernicus Open Access Hub platform (https://scihub.copernicus.eu/dhus/#/home) as well as on the Alaska Satellite Facility (https://www.asf.alaska.edu/).
The two images used for this exercise were
S1B_IW_GRDH_1SDV_20190203T082905_20190203T082930_014779_01B90C_150F, 1A level, ascending orbit, GRN product type and IW sensor mode.
1.1 Data preparation
This process must apply to both images.
Unzip the Sentinel-1 data in your working directory, Open SNAP software and call SAR images by clicking on File > Open Product and then selecting the manifest.safe file.
1.2 Apply Orbit File
This process must be applied to both images.
The orbit file provides an accurate position of SAR image and the update of the original metadata of SAR product, the orbit file is automatically downloaded from SNAP software. For this recommended practice, the "apply orbit file" procedure can be optional; nevertheless, it's highly recommended to ensure a successful spatial coregistration process.
Select Radar > Apply Orbit File, and then define all parameters according to Figure 2.
Select Radar > Radiometric > Calibrate, and set the “Processing Parameters”, select all polarizations and select Sigma0 as output band.
This process must be applied for both images.
1.4 Speckle Filtering
Select Radar > Speckle Filtering > Single Product Speckle Filter, and set the “Processing Parameters” as shown in Figure 5.
You can select another type of filter according with your expertise; however, we recommend using Lee and Frost filters as they are less degrading to the SAR image.
This process must be applied for both images.
Click on Radar > Geometric > Terrain Correction > Range-Doppler Terrain Correction, and define all “Processing Parameters” according to Figure 7.
This process must be applied for both image.
To get the same spatial subset to both images, click on View and select the following menu options: Statusbar, Synchronise Image Cursors and Synchronise Image Views.
Go to Raster > Subset, specify the area and parameters of the region of interest as shown in Figure 9.
This process must be applied to both images.
To ensure the same spatial coverage for both images, double click on any of the bands (vv or hv) of the second image and start again the “subset processing” as mention above.
Now, the new subset products must be converted to BEAM-DIMAP format. For this, right click over the name of each new subset product and select Save Product As, rename it if desired, click ok.
1.7 Coregistration stack
To generate a composite image from SAR data before and after the massive mudslide, a spatial coregistration procedure is necessary. For this, go to Radar > Corresgistration Stack > Tools > Create Stack, click the plus icon and select only the new subset products.
Define the rest of parameters as shown in Figure 14.
Open the new product as a composite image using the different polarizations. For this, right-click the name of the new product and select Open RGB Image Windows; choose the desired bands.
Example of the RGB image composite
Step 2: Processing
2.1 Change Detection
The Log Ratio is an algorithm used to the change detection procedure using mean ratio operator between two images of the same coverage area but taken at different times. To apply this procedure, click on Radar > SAR Applications > Change Detection, and define the processing parameters according to figure 17.
This procedure must be applied for all possible combinations between SAR images before and after the massive mudslide event. The following Log Ratio date combinations are required:
The massive mudslide and flooding area are identified by high-intensity pixels, identification is also possible by very low-intensity pixels. For this reason, PCA is used to regroup the pixel set optimally in order to achieve a digital enhancement of magnitude values associated with the affected area.
2.2 Coregistration stack (Log Ratio results)
To apply principal components analysis, a composite image from all Log Ratio results must be generated. For this, go to Radar > Corresgistration Stack > Tools > Create Stack, click on plus icon and select the four Log Ratio images. It is important to keep the order of the image combinations as listed in the previous step, "change detection" (Log Ratio):
Define the processing parameters according to Figure 21. Rename the file if desired.
2.3 Apply Principal Components Analysis (PCA)
Click on Raster > Image Analysis > Principal Components Analysis, and call the new product (Log Ratio stack), and select the four Log Ratio images. Define all “Processing Parameters” according to Figure 22.
The new product is created by four "out band components" and one "response band". As mention before, PCA implies a new data regrouping (a reversible orthogonal transformation), where towards the first "out band components" the variance is maximized. This means that the more significant pixel information is kept; while towards the last "out band component," all noise or redundant information is separated. The first three components can be used for the next steps. The “response band” represents each basis vector of the “out band components”, high values correspond to a better fit of data; so, this band can be used for the next steps as well.
“Out Band Components” results:
“Response band” result:
2.4 Supervise image classification
Each of the four final bands contains digital enhancement information about massive mudslide event and associated flooding. The differences in the spatial information provided by the final bands are linked to soil moisture conditions, differences between water bodies or by the large volume of sediments transported during a flood.
Thus the next step is to extract the affected area using a classification procedure. First, it is necessary to define the training polygons. Only two classes were defined: “no mudflow area” and “mudflow”; we recommend visualizing each of the final bands separately.
- double click on the final band selected
- go to New Vector Data Container and a new window will be opened
- name the polygon class "no mudflow area" and you can start to draw the polygon contouring
- if you desired to add another polygon of the same class "no Mudflow area", click on Polygon Drawing Tool and then start to draw it
- to add a new polygon class, go to New Vector Data Container and a new window will be opened
- name the new polygon class "mudflow" and click ok
- click over the image and a new window will open, the two classes are listed, select "mudflow"; now draw the polygon contouring
- if another polygon of the "no Mudflow area" class is needed, click on Polygon Drawing Tool and start drawing it
Go to Raster > Classification > Supervised Classification > Maximum Likelihood Classifier
Select the two “training vector” classes previously defined and the “feature bands” (out band components and response band) as shown in the Figure 29.
Supervised classification result:
2.5 Apply smooth filtering
To smooth the classified image result and regroup the pixels in an optimal way, a filtering process is needed. For this, go to Raster > Image Analysis > Texture Analysis > Grey Level Co-ocurrence Matrix, select only GLCM Variance option. Define all “Processing Parameters” according to Figure 31 and Figure 32.
Smooth filtering result:
Step 3: Post Processing
3.1 Getting a binary final image
After the smooth filtering procedure, the values of the final classified image convert from binary values to continuous raster cell values (stretch values); thus, converting a raster file to shape format can be complicated. An easy way to get a final binary image again is to apply a Supervised Classification a second time. For this, the last product from the “smooth filtering” procedure must be used to apply the Supervised Classification a second time.
3.2 Convert raster to shape file
Go to File > Export > GeoTIFF and select the last product (Supervised Classification a second time).
Open QGIS software and click on Raster > Conversion > Poligonize, and call the GeoTiff file.
The shape file is generated.
click the polygon that represents the spatial distribution of the massive mudslide; the polygone will change colour
go to the layers menu and right-click on the name of the polygone file and go to Export > Save As > Selected Features, save the new shape file as Esri Format
3.3 Visualization of results
Open Snap Software and call the RGB composite image generated in the procedure “Corregistration stack” (Step 1).
Go to File > Import > Vector Data, and call the new shape file as Esri Format. Thus, you can overlap the shape file associated to the affected flooding area with respect to the original SAR data.
The result of this recommended practice was compared with previous information provided by the International Charter Space and Major Disasters, where the massive mudslide was mapped using RapidEye high resolution data acquired after the event.
It is important to mention that the optical satellite image (RapidEye) used to identify the affected area has a higher spatial resolution than Sentinel-1. Despite this, the final results are very similar in spatial terms. This speaks about the effectiveness of our method of mapping this type of disaster event using Sentinel-1.