Supervised Classification for Wildfire Monitoring: Assessment of the Burnt Area


 Updated tutorial at 
https://fromgistors.blogspot.com/2017/01/wildfire-monitoring.html


Particularly after the summer, wildfires occur all over the world causing several damages and burning hectares of forest. Although sometimes fires are natural, it is important to monitor their environmental impacts.
Remote sensing can provide valuable information for monitoring wildfires. Watch the following footage by ESA Earth from Space (new videos are available every Friday) which illustrates a very interesting example of wildfire detected by satellite imagery.


This tutorial is about wildfire monitoring through the use of semi-automatic supervised classifications of remote sensing images.
We are going to classify a Landsat 8 image acquired over Alaska in 2013, during the Castle Rocks fire in Denali National Park and Preserve. Before starting the tutorial, please read this article about fires in Denali National Park and Preserve by NASA Landsat Science.
Study area, Alaska
(image available from the U.S. Geological Survey)

We are going to estimate the burnt area following these steps:
  1. Conversion of raster bands from DN to Reflectance;
  2. Creation of the ROIs and spectral signatures;
  3. Classification of the image using a threshold;
  4. Estimation of the burnt area.

The classification is focused on burnt surfaces; therefore we are going to identify only this class by using the Spectral Angle Mapping algorithm with the definition of a threshold.
Before starting this tutorial, if it is your first time using the Semi-Automatic Classification Plugin, please follow the basic tutorial here.
First, download the dataset from here that includes the Landsat image (available from the U.S. Geological Survey), and the metadata file. We are going to use the Landsat 8 bands (16 bit raster) described in the following table.

Band ID Landsat 8 Band Number Spectral Range
1 Band 2 Blue
2 Band 3 Green
3 Band 4 Red
4 Band 5 Near-Infrared
5 Band 6 Short Wavelength Infrared 1
6 Band 7 Short Wavelength Infrared 2

1. Conversion of raster bands from DN to Reflectance

The conversion of raster bands to surface reflectance, performing the image-based atmospheric correction using the DOS1 method, aims to improve the classification results.

Steps:
  • Open QGIS and start the Semi-Automatic Classification Plugin (SCP);
  • In the SCP Main interface select the tab Pre processing > Landsat;
  • Select the directory that contains the Landsat bands, and select the output directory where converted bands are saved;
  • Check the option Apply DOS1 atmospheric correction, and click Perform conversion to convert Landsat bands to reflectance;
SCP tab Landsat
  • At the end of the process, converted bands are loaded in QGIS; also, a virtual raster named landsat.vrt is loaded (containing all the Landsat bands converted to reflectance);
  • Select the Landsat virtual raster, left click and open its properties; in Style select band 4 (i.e. Near-Infrared) for the red band, band 3 (i.e. Red) for the green band, and band 2 (i.e. Green) for the blue band.
The result is shown in the following image.

Virtual raster of the Landsat image (RGB = 543)

2. Creation of the ROIs and spectral signatures

In this post we are going to use the Landsat virtual raster created before (one of the great features of GDAL).

Steps:
  • Select the Landsat virtual raster as input image; we must however define the band center wavelength in the Band set tab. As you can see, it is not possible to order the bands (they are already ordered), therefore we just need to select the Landsat 8 item under Quick wavelength settings;
SCP tab Band set
  • Define the training shapefile (for the ROI collection), and the signature list file (which stores the spectral signatures calculated from ROIs or imported from other sources) as described here (step 2);
  • In the dock ROI creation click the button + beside Create a ROI and then click a burnt area (brown/dark green pixels);
  • Under ROI Signature definition type a brief description of the ROI inside the field Class Information and Macroclass Information, and assign a Macroclass ID and Class ID;
  • In order to save the ROI to the training shapefile click the button Save ROI to shapefile; if the checkbox Add sig. list is checked, then the spectral signature is added to the Signature list table;
  • Define the class color with a double click on the Color column in the Signature list.
Creation of the first ROI and spectral signature
  • In the SCP Toolbar, select the Landsat virtual raster as input image; we must however define the band center wavelength in the Band set tab. As you can see, it is not possible to move the bands (they are already ordered), therefore we just need to select the Landsat 8 OLI item under Quick wavelength settings;
Selection of the Landsat 8 OLI settings

  • Define the training shapefile (for the ROI collection), and the signature list file (which stores the spectral signatures calculated from ROIs or imported from other sources) as described here (step 2);
  • In the dock ROI creation click the button + beside Create a ROI and then click a burnt area (brown/dark green pixels);
  • Under ROI Signature definition type a brief description of the ROI inside the field Class Information and Macroclass Information, and assign a Macroclass ID and Class ID;
  • In order to save the ROI to the training shapefile click the button Save ROI to shapefile; if the checkbox Add sig. list is checked, then the spectral signature is added to the Signature list table;
  • Define the class color with a double click on the Color column in the Signature list.
Repeat these steps in order to collect several ROIs.

3. Classification of the image using a threshold

SCP allows for classification previews that show very rapidly the classification results in small areas.
In this tutorial we need to set a threshold for the Spectral Angle Mapping algorithm in order to classify only pixels that are very similar to our spectral signatures, leaving unclassified the rest of the image.
If the angle between the pixel spectral signatures and the ones we collected in the previous step is below the threshold value, then pixels will be classified as burnt area.
What we are looking for is the maximum threshold value that allows for the identification of all the burnt areas in the image, without errors of commission (i.e. pixels incorrectly classified as burnt area).

Steps:
  • In the dock Classification, under Classification preview set Size = 500; check Use Macroclass ID, and select the Spectral Angle Mapping algorithm; under Threshold type 10 and click the button + and then click on the image; after a few seconds, the classification preview will be displayed; we can change the threshold value in order to improve the results;
Classification preview
  • In order to perform the final classification, under Threshold type 15 and click the button Perform classification and select where to save the output (e.g. classification.tif).
The final classification

You can download the Landsat classification, the ROIs, and the Signature file from here.

4. Estimation of the burnt area

SCP allows for the calculation of a classification report with the percentage and the area of land cover classes.

Steps:
  • Select the tab Post processing > Classification report of the SCP Main interface;
  • Select classification.tif beside Select the classification and check Use No data value leaving the value 0;
  • click Calculate classification report; after a few seconds the report will be displayed, showing the percentage and the area of the class.
SCP tab Classification report

These results indicate that about 218km2 were burnt in the study area.

However, the burnt area estimated in this tutorial is inaccurate for several reasons. For a more accurate estimation we would require more ROIs, considering the spectral similarity of burnt areas with certain type of soils. Also, field data is recommended for improving the classification process. Finally, it would be useful to classify an image before the fire, in order to assess the land cover change, and exclude previously burnt areas or bare soil from the computation.

The aim of this tutorial was to describe how to monitor the effects of wildfires estimating the burnt area in a very simple and rapid fashion. The spatial resolution of images (i.e. 30m) affects the area estimation, therefore we can assess only the effects of large fires.
Considering the free availability of Landsat images, monitoring land cover change at the regional scale can be both affordable and reliable.

The next week I am going to post a new tutorial about monitoring land cover change of forest, in particular caused by deforestation.

Please, remember that a Facebook group and a Google+ Community are available for sharing information and asking for help about the Semi-Automatic Classification Plugin.
Newer posts Older posts

Follow

RG