Geoscience Reference
In-Depth Information
Chapter 17
Case Study on Multispectral Land Cover
Classification
Land cover classification is one of the principal applications of remote sensing.
The goal is to distill satellite images into a spatially explicit information in the
form of a categorical land cover map. Image classification is considered to employ
a relatively standardized workflow. The process itself involves a wide variety of
geospatial operations, and consequently it is an ideal candidate for our case study.
Image classification can be divided into twomain types: unsupervised and supervised.
Unsupervised classification automatically clusters the image data and assigns classes
to each group. A supervised image classification relies on a training dataset to assign
class labels to previously unseen image pixels. The training dataset can be collected
in the field or digitized from existing reference maps and/or imagery. The assessment
of the classification accuracy of the map is generally compared to an independent
validation dataset.
The objective of this case study is to present an efficient workflow to process and
classify satellite data using a combination of command line utilities. The case study
is an example of how the different code snippets presented throughout this topic can
be merged into one processing chain to solve a real-world problem. As with most of
the examples presented in this topic, the case study can be replicated using the same
data and command line utilities, but it can equally be adapted to other datasets and
applications.
Here, we will assume the training and test datasets to be available as OGR vector
files. Training data are obtained from OpenStreetMap (OSM, see Appendix A.1.1).
An independent test set is obtained from the interpretation of high spatial resolution
imagery, using the OpenLayers plugin within QGIS, based on Bing and Google Earth
imagery. The input for the image classification is a georeferenced raster dataset in
GeoTIFF format, acquired over Ireland with the Landsat 8 sensor. Our focus here
is on a binary land cover classification problem that maps both a forest and non-
forest class. More specifically, we will classify tree cover. Without entering into the
semantics of what defines a forest, it can be assumed that tree cover is more closely
related to remote sensing observation data. The term forest is more commonly linked
to a land use class, which includes both stocked and unstocked forests. A multiclass
Search WWH ::
Custom Search