I’ve been asked for advice on processing Landsat data a few times in the last couple of weeks. So, here are a few tips.
Data can be acquired from a number of sources: http://www.Landsat.org has orthorectified data from a few time intervals; glovis.usgs.gov has a big archive of data (requires registration).
The United States Geological Service’s (USGS) Glovis offers more choice and a more convenient search tool. Glovis also has data from other sensors, such as Hyperion and ASTER.
Before we analyse the data we may want to preform some processing to normalise the data (to allow quantitative comparison between images) and remove atmospheric effects and noise. First we need to import the data: in ENVI we can use
Files>Open External Files>Landsat>Landsat GeoTiff with Metadata
Then we can start preparing the imagery (pre-processing). Calibration is the first step: here we can choose to convert the digital numbers to reflectance or radiance values. In most cases I’d choose reflectance. In ENVI there is an automatic Calibration tool:
Basic Tools>Pre-processing>Calibration Utilities>Landsat Calibration
Next, we want to remove the effects of atmospheric scattering. Some light is scattered by water vapour and aerosols, particularly at lower wavelengths in the blue part of the Electro-magnetic spectrum. If you have more comprehensive tools, such as the ENVI Atmospheric Correction module, you can perform corrections based on radiative transfer modelling. Otherwise we can do one of two things: assume that there will be no atmospheric effects (e.g. over deserts and other arid regions); or we can use an empirical correction. A common method is the Dark Object Subtraction.
Dark Object Subtraction (DOS) assumes that reflectance from dark objects includes a substantial component of atmospheric scattering. Hence we can measure the reflectance from a dark object, such as a deep lake, and subtract that value from the image. DOS in ENVI is found:
Basic Tools>Preprocessing>General Purpose Utilities>Dark Subtract
If your image contains noise, such as striping or detector effects, we can attempt to remove it in number of ways. Destriping removal tools are available. These normally assume striping to be regular and are based on identifying the fequency of strips. Alternatively we can mask out stripes if they have a distinctive signature, such as 0 or -9999. Masking however only excludes these pixels from analysis. A third option is to use a method such as Principal Components Analysis (PCA). For Landsat-7 ETM+ imagery acquired after the SLC failure in 2003 striping will be a problem.
Principal Components Analysis transforms the image data into a set of uncorrelated variables using statistical methods. The result of PCA is a image dataset in which each band is uncorrelated with the other bands: each band shows unique information. In normal image data neighbouring spectral bands are strongly correlated and therefore contain redundant information. By using PCA we can reduce this redundancy and emphasise the uncorrelated variables which are ‘hidden’ beneath the correlated data. What is more, PCA often reveals systematic noise in the data, which we can then remove. Incidently, PCA is used in hyperspectral remote sensing to reduce the number of bands used in image analysis without reducing the information content by reducing redundancy.
Figure: the first Principal Component (left) and 20th PC band from a Hyperion image subset over northern Finland. The 20th PC is almost entirely instrument noise.
PCA is a three step process: forward rotation, analysis, inverse rotation. The forward rotation calculates the principal components. These are then viewed and redundant bands are identified (i.e. those which comprise of noise). In the inverse rotation the original dataset is recreated without redundant information. PCA in ENVI is found here:
Transform>Principal Components>Forward PC Rotation
Transform>Principal Components>Inverse PC Rotation
The forward PC bands can also be used in their own right, for example as input to a classification. However, it should be remembered that the PCA bands are not a representation of spectral reflectance but uncorrelated information in multi-spectral imagery.
The pre-processing prepares the satellite imagery for analysis. The aim is to normalise the imagery to allow inter-comparison between scenes: for example, for time-series analysis. Correcting for atmospheric effects, reducing noise, and converting the data to a geophysical unit such as reflectances, achieves this. The data are now ready for analysis.