From GIS to Remote SensingA supervised classification of remote sensing images is a processing technique that allows for the identification of materials in the image, according to their spectral signatures (see here for further definitions about remote sensing).
The main advantage of this approach is that an entire image can be processed rapidly, producing the land cover classification thereof.
This post is about the interpretation of remote sensing images that is a fundamental phase of the ROI creation, which is a required step for the semi-automatic classification.
In this previous post, I have described the land cover classification of a Landsat image, using the Semi-Automatic Classification Plugin for QGIS. The following images illustrate the input image and the resulting land cover classification.
 |
| A Landsat 8 image (available from the U.S. Geological Survey) displayed in QGIS with color composite RGB=543 |
 |
| The land cover classification of the Landsat image, using the Semi-Automatic Classification Plugin for QGIS |
The semi-automatic classification of remote sensing images requires some user inputs, which are the Regions of Interest (ROIs). ROIs are groups of homogeneous pixels that are identified in the image, which define a land cover class, as illustrated in the following example.
ROIs are used by the program in order to assess the spectral characteristics thereof and therefore to assign a land cover class to each image pixel according to a classification algorithm. The spectral resolution of images (i.e. the number of bands) is an important parameter for semi-automatic classifications, because it determines the level of precision for the identification of materials (i.e. the more are the spectral bands, the more classes can be identified by their spectral signatures).
The identification of pixels in the image belonging to a certain land cover class is a fundamental phase, in order to assign the correct class to each ROI. This operation requires the user to know what are the objectives of the land cover classification (e.g. forest monitoring, urban sprawl assessment, crop identification, etc.), and which are the image resolutions (in particular, spatial and spectral resolutions).
Therefore, image interpretation is an essential skill for any land cover classification. Several tips for the interpretation of satellite images are provided in this article by the NASA Earth Observatory, which I invite you to read. The article (by Holli Riebeek and designed by Robert Simmon) provides several examples about the following topics:
 |
| In yellow a ROI of water class identified over a lake (i.e. the black area) |
ROIs are used by the program in order to assess the spectral characteristics thereof and therefore to assign a land cover class to each image pixel according to a classification algorithm. The spectral resolution of images (i.e. the number of bands) is an important parameter for semi-automatic classifications, because it determines the level of precision for the identification of materials (i.e. the more are the spectral bands, the more classes can be identified by their spectral signatures).
The identification of pixels in the image belonging to a certain land cover class is a fundamental phase, in order to assign the correct class to each ROI. This operation requires the user to know what are the objectives of the land cover classification (e.g. forest monitoring, urban sprawl assessment, crop identification, etc.), and which are the image resolutions (in particular, spatial and spectral resolutions).
Therefore, image interpretation is an essential skill for any land cover classification. Several tips for the interpretation of satellite images are provided in this article by the NASA Earth Observatory, which I invite you to read. The article (by Holli Riebeek and designed by Robert Simmon) provides several examples about the following topics:
- Identification of the image scale, which is related to the spatial resolution; the pixel size determines the level of detail of the objects identified in a classification;
- Identification of patterns and shapes in the image, which is very useful for the correct creation of ROI polygons;
- Interpretation of an image in true or false colours, which can highlight certain materials; for example, the previous image illustrates a Landsat image with the color composite RGB=543 (Near-Infrared, Red, and Green bands), which is useful for the identification of vegetation (displayed in red);
- Identification of clouds and shadows in the image; remote sensing images are often affected by clouds and shadows, which must be considered properly during the image preprocessing in order to avoid classification errors (in a future post I am going to describe how to avoid such errors);
- The importance of the correct geographic position of the image, which is related to the georeferencing process; in particular, the position error should always be lower than pixel size;
- The importance of prior knowledge of the place, for the correct interpretation of an image; for example, it is useful to have additional information from field survey in order to improve the classification results.
Finally, I would like to inform you about the availability of other useful information provided by the NASA; in particular, the following articles about using Landsat images for mapping, and about the dimensions of remote sensing.