An MLP band reducer (predict a trained network across the raster).

Returns a reducer function fn(x, dims) for reduce_over(cube, fn, over = "band"): standardises the band vector (optional), then applies dense layers act(W_l x + b_l) with ReLU between layers and output_activation on the last, yielding one value per pixel. Weights come in as plain R matrices – e.g. torch nn_linear layers, whose $weight is already (n_out, n_in) so as.matrix(layer$weight) / as.numeric(layer$bias) drop straight in. Dropout layers are inference-time identities: skip them.

Usage

mlp_project(
  weights,
  biases,
  center = NULL,
  scale = NULL,
  output_activation = c("identity", "sigmoid"),
  qa_plane = NULL,
  qa_floor = NULL
)

Arguments

weights

List of layer weight matrices, each (n_out, n_in), applied in order; n_in of the first layer = number of bands.

biases

List of bias vectors (n_out each), same length as weights.

center, scale

Optional per-band standardisation applied first: (x - center) / scale. Length = number of bands.

output_activation

"identity" or "sigmoid" on the final layer (hidden layers are ReLU).

qa_plane

Optional 1-based index of a QA plane riding as the LAST plane of the input cube (must equal n_in + 1). Predictions where the QA value is nodata (or below qa_floor) are NaN. Carrying QA inside the cube lets the whole prediction run as a single read with no separate masking pass.

qa_floor

Optional minimum QA value; below it the prediction is NaN. Only used with qa_plane.

Value

A function fn(x, dims) suitable for reduce_over() over = "band".

Details

NaN in any feature band yields NaN for that pixel (complete-cases semantics); gate QA upstream by mapping bad pixels' features to NaN. Target back-transforms (expm1, sinh, ...) compose downstream as an ordinary lazy_map.

See also

band_project() for the linear case, reduce_over()