Compatible Estimators

sklearn-raster can wrap estimators from scikit-learn and compatible third-party packages to create a FeatureArrayEstimator that supports raster-based predictions. To be compatible, estimators should implement a fit method and one or more output methods like predict, predict_proba, transform, or kneighbors which accept an array of features and an optional array of targets (for supervised learning), and return one or more corresponding arrays. The wrapper extends these methods to accept and return rasters, i.e. arrays of pixels with spatial and/or temporal dimensions.

Supported Features

sklearn-raster does support:

• Supervised and unsupervised estimators
• Regressors, classifiers, clusterers, and transformers
• Single-output or multi-output predictions
• Output methods that return tuples of arrays, e.g. kneighbors
• Third party estimators like xgboost and sknnr

Unsupported Features

There are a few unlikely caveats that would make an estimator incompatible. sklearn-raster does not support:

• Estimators that return an unpredictable number of output targets, such as a predict method that returns fewer targets than the estimator was fit with. To enable lazy computation, sklearn-raster must be able to predict output shape from the input data.
• Estimators that modify the output shape, i.e. that does not return exactly one output sample for every input sample. In this case, sklearn-raster isn't able to infer which input coordinate corresponds to which output coordinate, and methods will fail unpredictably.
• predict_proba methods on multi-output classifiers (see #56).