An image segmentation approach can provide an interesting analytical framework for remote sensing classification problems as well as dealing with some nuisance aspects of hyper-variable data (eg., very high resolution). The idea behind object-oriented classification is to perform, what is commonly referred to as, an image segmentation with the intent of producing image objects (polygons) representing homogeneous regions of multivariate information in the image stack. This can include ancillary data such as vegetation (eg., NDVI) and textural (eg., grey-level co-occurrence or STVD of NIR) metrics. Some approaches, such a eCognition, include a measure of contrast and texture in their algorithm negating the need for textural data so, be aware of the statistic you are using in segmenting the image.
Regardless of the algorithm applied, the end result should be a set of image objects that minimize within unit variance and maximize between unit variance. The interesting aspect of these types of approaches is that they can be hierarchical, with various levels of generalization created based on varying the segmentation model parameters. In simplest terms you could create something like a 1st level of forests and non-forest with a 2nd level (finer-grained polygons) representing forest species composition.
In using image objects for classification, you are functionally smoothing the variation to better represent the patterns associated with your modeled process. Very high resolution imagery can be very problematic for certain statistics, due to the autocorrelation/lack of independence, and the pixel-level variation does not necessarily match the classification schema. The end result often produces an estimate that better represents the pattern of your process and reduces error which can be quite notable in pixel-level validations.
A common analytical workflow for this type of classification would be:
- Define your hierarchy and associated classification schema
- Apply a segmentation algorithm generating image objects for each
level in the hierarchy
- Collect band statistics for the image objects. This can be done
using a zonal statistics approach and should include statistics that
represent the central tendency (eg., mean, median) and variation
(eg., variance, standard deviation, MAD) for each band.
- Assign the band statistics to the training data that defines the
- Apply an appropriate statistical model, for classifying the image
objects, utilizing the training data. This model can then be
estimated to the band summaries for the image objects.
Stepping into a statistical software for this they of analysis, once you have the image objects created and summarized, opens many possibilities in available models as well as providing a robust validation and simulation platform. A fairly simple model to implement for this type of problem, in R, would be Random Forests. There is very clear methodology for building the classification model, validation and predicting it to a set of image-object rasters or vector polygons. Here is a very basic example.