Remote Sensing Image Analysis

Some of the most basic image analysis you can do is classification of land cover types. In this case the first step would be to enhance your image using a variety of options, then run a classification either unsupervised (no ground control points) or supervised (using human defined spectral signatures using areas of interest and adding them to the signature editor).

Landsat Imagery analysis can do a lot more than just classify landcover types. Examples of additional methods include Fourier Analysis, Principal Component Analysis (PCA) and colour space transformations. This allows the extraction for more information from an image.

The purpose of this project is to learn three new methods of image analysis including Principal Components Analysis (PCA), Tasseled Cap Transformation, and Normalized Difference Vegetation Index (NDVI) on a Landsat scene. Landsat scenes can be downloaded from earthexplorer.usgs.gov/.

If you have a version of ERDAS Imagine and would like to follow this workflow I would recommend downloading imagery in the GeoTIFF file format. This will give you one TIFF file for each spectral band. The imagery should already be orthorectified.

Top-Of Atmosphere Reflectance

Basic classification and image enhancement use DN Values. These can be analyzed and used for spectral signatures. This represents the amount of reflected energy but does not consider incident energy (variable with sun angle and time of year)

The conversion of DN values to TOA reflectance values can help with analysis especially when analyzing vegetation indices.

This project will use TOA reflectance which considers seasonal sun angle variations. More complete sun information can be found using an ATCOR3 atmospheric correction (PCI Geomatics).

During the winter months when the earth is tilted away from the sun, the incident radiation is of lower intensity. The TOA reflectance accounts for this change.

Converting DN to TOA Reflectance

First you must import the landsat 8 data. You can do this in ERDAS or PCI. To convert from the provided (by Landsat 8 USGS) DN’s to the rescaled TOA reflectance, a rescaling coefficient is provided in the metadata file. This also has information on the thermal constants to convert TIRS to brightness temperature.

Although Landsat satellites image from nadir, the sun is coming in from a tangent (solar elevation angle or sun-elevation) and this needs to be considered.

This can be done in both ERDAS and PCI. I do not have the proper ERDAS extension so the Spatial Modeller was used to calculate the math manually. An example is shown in the tutorial. I did not have the best of luck with this so I redid the atmospheric corrections in PCI. The steps for this are shown here.

 

  • Raw Imagery

    Raw Imagery

  • TOA Corrected

    TOA Corrected

The slider about demonstrates that the TOA correction did not make a large difference in the imagery. Using an ATCOR correction or an imagery with more haze issues may result in a larger difference between the two images. Another factor that affects the ability of the correction is the wide scope in landcover type. The mountains create have cloud cover, snow cover and shadowing affects. This is difficult the bright and dark responses of these features skew the spectral range. An area with only one main type of landcover may see more variable results.

Principal Components Analysis

The principal components analysis (PCA) is a method that condenses information into intercorrelated variables by collecting the majority of variance into the first band. The remaining variance is assigned to additional bands depending on how many bands the user inputs. This removes redundancy and decreases file size.

http://wiki.landscapetoolbox.org/doku.php/remote_sensing_methods:principal_components_analysis

Although this may not seem important when you are dealing with only 3-7 bands, as the band number increases on more recent satellites, this can be a huge time saver. For example, imagine that you are dealing with hyperspectral imagery that includes 256 bands. Reducing the number of bands can be very beneficial for analysis.

The results of this analysis are shown below. More information on completing this step can be found here.

  • Raw Imagery

    Raw Imagery

  • Principal Component Analysis

    Principal Component Analysis

The principal component analysis (PCA) was very useful in this image as it helped remove a lot of the dark shadowing affects caused by terrain in the mountains to the west. It’s strengths are in determining the locations of high urban density and abnormal spectral signatures (such as clouds and snow which appear bright red in this band combination).

Additionally this analysis method is also very good at distinguishing bare ground from vegetation although it does not recognize vegetation health or growth stage.

This method is fast and requires less processing time than many other types of imaging. Unlike supervised classification it does not require much user input although to reduce your file size as much as possible it is advisable to consider the eigenvalues and only output as many principal components as necessary. The table created in this project is shown below. Three principle components were used as >99% of the of the results fell within these three dimensions.

Tesseled Caps Transformation

The Tesseled Cap Transformation makes three components that are related to brightness, greenness and wetness.

The tasseled-cap transform takes a linear combination of the original image bands which is useful for vegation mapping. Each new band is created by summing image band 1 x a constant + imaage band 2 x a constant for example. For more information select “view” when running the tool and it will show you the modeller.

Step by Step

  • Raw Imagery

    Raw Imagery

  • Tesseled-Cap Transformation

    Tesseled-Cap Transformation

TCT outputs as many bands as are input but in a landsat image the output bands 4 and 6 are often just noise.

Band 1 is brightness, band 2 is greenness and band 3 is wetness. Depending on the combination you can focus on getting results for any of these three variables.

These bands are difficult to analyze when displayed in combination but they can be experimented with to help distinguish wetness, brightness and greenness of the imagery.

In this example, TCT is good at distinguishing water, urban areas and vegetation. This can be used to help determine urban development, landuse change and ground surface moisture contents.

Analysis of this enhancement seems more complex than PCA and NDVI and would require further study.

Normalized Difference Vegetation Index

The main use of NDVI classification is vegetation productivity. By analyzing reflectance and absorption plant growth and health can be determined. This uses the equation (NIR-Red)/(NIR+Red).

The values range from -1 to 1. Values around zero are typically barren ground (rock, sand, snow). Low positive values are shrub and grassland and high positive values are quickly growing productive plants. Negative values correspond to water.

After the NDVI raster is created you can open it in psuedocolour and change the colours to suit your needs as shown below.

 

  • Raw Imagery

    Raw Imagery

  • NDVI

    NDVI

  • NDVI Psuedo Colour Classified

    NDVI Psuedo Colour Classified

    White – new/productive growth, Dark green – vegetation, Brown – bare ground, Aqua – Moist ground, urban and shadowed vegetation, Blue – Lakes

The imagery above does a really great job at distinguishing new growth from old growth and areas with a high production value. It distinguishes vegetation within the mountains and can help determine areas of the city with more growth. 

 

Conclusions

The TOA did not work using the ERDAS modeller. This did work in PCI although it was not used for the subsequent enhancements. Each of the different enhancement provides variable results and is useful for determining different features of the landscape such as new growth (NDVI), bare soil (PCA) and urban environments (TCT and PCA). TCT was very useful for determining soil and vegetation moisture as well as distinguishing urban landscapes from bare soil.