Classifier Options

Reference > Classifier Options

The Classifier Options Datasheet sets the model and how many pixels it learns from. It is used by 1-Train Classifier and read again by 2-Predict.

Field Description
Sample size Total number of pixels sampled from the training rasters. Defaults to 10000. Pixels are sampled across a grid of spatial blocks so the samples are spread out, and the presence class is kept at 15–35% of the sample when it is rare. 80% of the blocks are used for training and 20% for testing, so the evaluation metrics come from areas the model did not see. Larger samples usually improve the model but take longer to train.
Model type The classifier. Random Forest (default; the ranger package) is a robust general choice and reports variable importance. MaxEnt (the maxnet package) is a presence/background model often used for species and habitat mapping. CNN (the torch package) is a convolutional neural network that can learn more complex patterns but needs more data and time.

See Advanced Classifier Options for pre-processing, tuning and threshold settings.