How to generalize training data for a random forest classifier?

Question

I am working with a random forest classifier in eCognition (new with eCognition 9.0). The image shows 8 NAIP tiles I am attempting to classify using approximately 100 training points. The training data (blue points) are very simple and are comprised of a shapefile indicating the location and class of vegetation. I am classifying the images on a tile by tile basis, so I am looking for a way to utilize all of the training data rather than only the points that lie within the extent of an individual tile.

What methods are available to utilize the full range of training data on a tile by tile basis? I have considered extracting the pixel values for each point and exporting that as a table, although I have not found a way to use tabular data to train the random forest classifier in eCognition.

Answer 1

Why keep the entire analysis in eCognition? Once you have your image objects derived, export them and run the model in R. You have far more control of the model in R (e.g., model specification, multi-colinearty test, model selection, etc...) and there is no problem fitting a model to all of the data and predicting it to subsets represented by the tiles.

I would create a workflow where:

1) image objects are created and summarized in eCognition.

2) export results as image object polygons or rasters representing spectral/textural statistics of image objects.

3) read point training data and image objects, as polygon SpatialPolygonsDataFrame or raster stack object, into R.

4) assign image object statistics to training points. Due to the size of the problem, steps 3 & 4 may need to be done iteratively to construct data utilized in final model.

5) fit random forests model, applying multi-colinearity test and model selection.

6) predict global RF model to each tile of data.

Answer 2

This is not the exact answer but could be used as a workaround.

I guess that you are not using the 8 tiles together for memory reason, but your area seems to be quite homogeneous. So you could degrade the resolution of your images (e.g. with a factor 2 or 3) and create a mosaic. Then you train your classifier on the mosaic image and you "save to file" the scene variable containing your model parameters. This file can be uploaded as a scene variable in another project, and you can use it to apply the classifier with each tile.

Answer 3

There is a sample ruleset on the eCognition Community Ruleset Exchange that shows a work-around. Note that you may need to register to access the ruleset exchange link. The rule-set description states the following:

This zip archive contains example data to train and apply the classifier algorithm on multiple scenes. At the moment you can do this with a single project which contains the scenes as separate maps. This is a workaround if you want to train a classifier not over unlimited number of scenes. eCognition will cover the use case, to train a classifier automatic over a unlimited number of projects, hopefully in one of the next versions.

You can use the same approach also for SVM, Bayes and kNN.