PredictionRequests

The prediction_requests key of the config.

Controls the prediction requests for the model.

Field	Type	Required	Default	Description
partitioners	Partitioners	yes		Controls the partitioning of the input data into smaller units.
postprocessors	Postprocessors	no	null	Controls the postprocessing of the prediction results and their assembly into final outputs.

Partitioners

Field	Type	Required	Default	Description
request_to_partitions	Partitioner	yes		Controls the partitioning of the initial request into smaller units for machine-level parallelism.
partition_to_windows	Partitioner	yes		Controls the subdivision of a request partition into smaller units for process-level parallelism.

Postprocessors

Post-processing configuration.

Field	Type	Required	Default	Description
additional_steps	array of AdditionalPostprocessor	no	null

Partitioner

One of the following, selected by its name field:

name	Type
fixed_window	FixedWindowPartitioner
grid	GridPartitioner
manual	EscapeHatchComponent
noop	NoopPartitioner

AdditionalPostprocessor

One of the following, selected by its name field:

name	Type
value_transform	ValueTransformPostprocessor
vectorize	VectorizePostprocessor

NoopPartitioner

Partitioner that does nothing.

Field	Type	Required	Default	Description
name	"noop"	yes

GridPartitioner

Partitioner that uses a grid to partition the input data.

Field	Type	Required	Default	Description
name	"grid"	yes
grid_size	number (> 0)	yes		The size of the grid cells in units of the output projection.
overlap_size	number (≥ 0)	no	0	The size of the overlap between adjacent grid cells in units of the output projection.
projection	Projection	no	null	Projection strategy to use for output geometries. If not provided, no re-projection will take place.

FixedWindowPartitioner

Partitioner that creates fixed-size square windows.

Window size is determined by the window_sizing strategy. Defaults to MatchInputSize, which uses input_preprocessing.default.input_size.

Field	Type	Required	Default	Description
name	"fixed_window"	yes
window_sizing	PartitionWindowSizing	no	{"name": "match_input_size"}	Strategy for determining the window size.

VectorizePostprocessor

Polygonize a raster prediction into GeoJSON vector features.

Field	Type	Required	Default	Description
name	"vectorize"	yes
skip_label_ids	array of integer	no	null	Label IDs to exclude from vector output.
sieve_size	integer (≥ 0)	no	5	Minimum connected-pixel group size; smaller groups are removed before polygonizing. Higher values drop more small features, reducing noise but also dropping small real detections.
simplify_tolerance_px	number (≥ 0.0)	no	1.0	Douglas-Peucker tolerance in pixels of the output resolution, applied in the native CRS before reprojection. 0 disables; sub-pixel values do approximately nothing (vertices sit on the pixel grid).
coordinate_precision	integer (≥ 0)	no	5	Decimal places to round output coordinates (5 ≈ 1 m precision).
max_features	integer (> 0)	no	null	Hard cap on output feature count; keeps N largest by area when hit.
field	string	no	null	Name of the output field to polygonize as class IDs. Its band and legend come from its declaration in data.output.fields. Defaults to the only field when there is one.
auxiliary_fields	map of string to ModeAggregation or MinAggregation or MaxAggregation or PercentileAggregation	no	null	Other output fields to sample onto each polygon, keyed by field name. The field name becomes the property name on the output feature. Segmentation fields are written as their labels and date fields as ISO dates (YYYY-MM-DD).

ValueTransformPostprocessor

Transform pixel values of raster prediction outputs.

The transform is applied exactly once: either to each window's raster before combining (window, the default) or to the final combined raster (request). Window level casts to smaller dtypes early, so combining moves less data.

Embedding quantization is not configured here -- the runner power-quantizes embeddings to int8 automatically.

Field	Type	Required	Default	Description
name	"value_transform"	yes
level	"window" or "request"	no	"window"
transform	LinearValueTransformConfig	yes		The value transform to apply.

Projection

One of the following, selected by its method field:

method	Type
crs	CRSProjection
use_utm	UseUTMProjection

PartitionWindowSizing

One of the following, selected by its name field:

name	Type
fixed_size	FixedSize
match_input_size	MatchInputSize

ModeAggregation

The most common value under the polygon, for categorical fields.

Field	Type	Required	Default	Description
aggregation	"mode"	no	"mode"

MinAggregation

The smallest value under the polygon, such as the earliest date. One outlier pixel sets it.

Field	Type	Required	Default	Description
aggregation	"min"	yes

MaxAggregation

The largest value under the polygon, such as the latest date. One outlier pixel sets it.

Field	Type	Required	Default	Description
aggregation	"max"	yes

PercentileAggregation

A percentile of the values under the polygon, by nearest rank, so it is always a value the band holds.

For ordered fields such as dates: a low percentile is the robust form of min (10 tracks a change's onset while ignoring outlier pixels), a high one the robust form of max.

Field	Type	Required	Default	Description
aggregation	"percentile"	yes
percentile	number (≥ 0, ≤ 100)	yes

LinearValueTransformConfig

Linear transform: output = scale * input + offset.

Field	Type	Required	Default	Description
type	"linear"	no	"linear"
scale	number	no	1.0
offset	number	no	0.0
output_dtype	"uint8" or "int16" or "int32" or "float32"	no	null	The output dtype to transform to. If not specified, the output dtype will be the same as the input dtype.

UseUTMProjection

Field	Type	Required	Default	Description
method	"use_utm"	yes

CRSProjection

Field	Type	Required	Default	Description
method	"crs"	yes
crs	string	yes		The CRS code of the output projection.
x_resolution	number (> 0)	yes		The x resolution of the output projection, in projection units.
y_resolution	number (< 0)	yes		The y resolution of the output projection, in projection units.