Model

The model key of the config.

Field	Type	Required	Default	Description
encoder	Encoder	yes		The encoder to use for the model. This produces embeddings from preprocessed modality data for use as input to task heads.
tasks	map of string to ModelTask	yes		The tasks to perform with the embeddings produced by the encoder layer. Each task will produce predictions for the original inputs.

Encoder

One of the following, selected by its name field:

name	Type
manual	ManualEncoder
olmoearth	OlmoEarthEncoder

ModelTask

One of the following, selected by its name field:

name	Type
classification	ClassificationTask
detection	DetectionTask
embedding	EmbeddingTask
manual	ManualTask
per_pixel_regression	PerPixelRegressionTask
regression	RegressionTask
segmentation	SegmentationTask

OlmoEarthEncoder

OlmoEarth encoder configuration.

Field	Type	Required	Default	Description
name	"olmoearth"	yes
patch_size	integer (≥ 1, ≤ 8)	yes		The size of the patch to use for the model
source	OlmoEarthSource	yes		Where to load the OlmoEarth encoder from
output	OlmoEarthEncoderOutput	yes		Which of the model's outputs to use as the embedding.
use_legacy_timestamps	boolean	yes		Feed the model dummy monthly timestamps instead of the imagery's capture times.
token_pooling	boolean	no	true	Pool each patch's tokens over timesteps and band sets into one embedding. Set false to keep a token per timestep, for manual-task decoders that read the time series (e.g. a breakpoint scan).
location_query_patch_size	integer (≥ 1)	no	null	Pixels per output embedding along each side, for encoders with an AttentionPooler (OlmoEarth v1.3+). Must divide patch_size; e.g. 1 gives one embedding per input pixel from patch_size 2 tokens. If omitted, the output is one embedding per patch.

ManualEncoder

Escape hatch encoder providing raw rslearn encoder layers, e.g. OlmoEarth wrapped in SimpleTimeSeries.

Inputs are still prepared the OlmoEarth way (band order and normalization), so the layers must accept what the OlmoEarth encoder accepts. The transpiler can't look inside the layers, so the two properties the decoders depend on are declared here. A wrapped OlmoEarth loads its weights from its own init_args (a Hugging Face model_id or a path); registry foundation models aren't resolved for it.

Field	Type	Required	Default	Description
name	"manual"	yes
encoder_layers	array of ComponentConfig (min 1 items)	yes		The rslearn encoder layers, passed through to the model config as-is.
patch_size	integer (≥ 1)	yes		The factor by which the encoder's feature map is smaller than its input, e.g. the wrapped OlmoEarth patch size.
embedding_size	integer (≥ 1)	yes		The number of channels in the feature map the encoder hands to the decoders. A wrapper that concatenates features across groups multiplies the wrapped encoder's embedding size by the number of groups.

SegmentationTask

Field	Type	Required	Default	Description
name	"segmentation"	yes
decoder	SegmentationDecoder	yes

PerPixelRegressionTask

Field	Type	Required	Default	Description
name	"per_pixel_regression"	yes
decoder	PerPixelRegressionDecoder	yes

ClassificationTask

Field	Type	Required	Default	Description
name	"classification"	yes
decoder	ClassificationDecoder	yes

RegressionTask

Field	Type	Required	Default	Description
name	"regression"	yes
decoder	RegressionDecoder	yes

DetectionTask

Field	Type	Required	Default	Description
name	"detection"	yes
decoder	DetectionDecoder	yes
box_side_length	integer	no	null	When set, each annotation's geometry is replaced by a fixed box of this side length (in pixels) centered on the geometry's centroid, and that box becomes the training label. When unset, the geometry's native bounding box is used. Required for point annotations (which have degenerate bounding boxes); optional for polygons. Must be less than input_size.

EmbeddingTask

Prediction-only task that produces encoder embeddings directly.

No decoder or training configuration -- the encoder's feature map is saved as the output via EmbeddingHead.

Field	Type	Required	Default	Description
name	"embedding"	yes

ManualTask

Escape hatch task providing raw rslearn decoder_layers and task config.

Field	Type	Required	Default	Description
name	"manual"	yes
decoder_layers	array of ComponentConfig	yes
task	ComponentConfig	yes

OlmoEarthSource

One of the following, selected by its name field:

name	Type
distributed_checkpoint	OlmoEarthFromDistributedCheckpoint
huggingface	OlmoEarthFromHuggingFace
model_path	OlmoEarthFromModelPath
registry	OlmoEarthFromRegistry

OlmoEarthEncoderOutput

Allowed values: "patch_tokens", "student_location_tokens"

ComponentConfig

A reference to an arbitrary class to be initialized: class_path + optional init_args.

Field	Type	Required	Default	Description
class_path	string	yes
init_args	object with arbitrary keys	no	null

SegmentationDecoder

One of the following, selected by its name field:

name	Type
unet	UnetSegmentationDecoder
upsample_conv	UpsampleConvSegmentationDecoder

PerPixelRegressionDecoder

One of the following, selected by its name field:

name	Type
unet	UnetPerPixelRegressionDecoder

ClassificationDecoder

One of the following, selected by its name field:

name	Type
pooling	PoolingClassificationDecoder

RegressionDecoder

One of the following, selected by its name field:

name	Type
pooling	PoolingRegressionDecoder

DetectionDecoder

One of the following, selected by its name field:

name	Type
faster_rcnn	FasterRCNNDetectionDecoder

OlmoEarthFromHuggingFace

OlmoEarth encoder loaded from HuggingFace model ID.

Field	Type	Required	Default	Description
name	"huggingface"	yes
model_id	string	yes		The ID of the model to use (sourced from HF)
embedding_size	integer	no	null	The number of dimensions in the embeddings the encoder produces. If omitted, the size listed for model_id in rslearn's EMBEDDING_SIZES is used. Required when output is student_location_tokens.

OlmoEarthFromRegistry

OlmoEarth encoder loaded from OlmoEarth's foundation model registry.

Field	Type	Required	Default	Description
name	"registry"	yes
model_name	string	yes		The name of the foundation model to use from OlmoEarth's registry
embedding_size	integer	yes		The number of dimensions in the embeddings the encoder produces.

OlmoEarthFromModelPath

OlmoEarth encoder loaded from a model checkpoint + config path. For unpublished models. Note: the embedding size must be explicitly provided, as we have no catalog or metadata to consult.

Field	Type	Required	Default	Description
name	"model_path"	yes
model_path	string	yes		The path to the model checkpoint + config to use
embedding_size	integer	yes		The number of dimensions in the embeddings the encoder produces.

OlmoEarthFromDistributedCheckpoint

OlmoEarth encoder loaded from distributed checkpoint files + config. For in-flight pretraining checkpoints. Note: the embedding size must be explicitly provided, as we have no catalog or metadata to consult.

Field	Type	Required	Default	Description
name	"distributed_checkpoint"	yes
checkpoint_path	string	yes		The path to the distributed checkpoint files + config to use
embedding_size	integer	yes		The number of dimensions in the embeddings the encoder produces.

UpsampleConvSegmentationDecoder

A decoder that upsamples the embeddings using a convolutional layer before passing to a segmentation decoder head.

Field	Type	Required	Default	Description
name	"upsample_conv"	yes

UnetSegmentationDecoder

A decoder that uses a U-Net architecture to perform segmentation.

Field	Type	Required	Default	Description
name	"unet"	yes
conv_layers_per_resolution	integer (≥ 1)	no	1	Number of convolutional layers at each resolution level.
num_channels	map of string to integer	no	null	Explicit channel counts per resolution level (e.g. {8: 512, 4: 512, 2: 256, 1: 128}). When None, rslearn auto-calculates.

UnetPerPixelRegressionDecoder

A decoder that uses a U-Net architecture with a per-pixel regression head (single channel output).

Field	Type	Required	Default	Description
name	"unet"	yes
conv_layers_per_resolution	integer (≥ 1)	no	1	Number of convolutional layers at each resolution level.
num_channels	map of string to integer	no	null	Explicit channel counts per resolution level. When None, rslearn auto-calculates.

PoolingClassificationDecoder

A decoder that uses global spatial pooling over encoder features followed by fully-connected layers and a classification head. Used for window-level (vector) classification tasks.

Field	Type	Required	Default	Description
name	"pooling"	yes

PoolingRegressionDecoder

A decoder that uses global spatial pooling over encoder features followed by fully-connected layers and a regression head. Used for window-level (vector) regression tasks.

Field	Type	Required	Default	Description
name	"pooling"	yes

FasterRCNNDetectionDecoder

A decoder that uses a Faster R-CNN architecture to perform detection + classification.

Field	Type	Required	Default	Description
name	"faster_rcnn"	yes
anchor_sizes	array of array of integer	yes		Anchor box sizes in input image pixels for the Region Proposal Network. Each inner list corresponds to one encoder feature map level. OlmoEarth produces a single feature map, so use one inner list, e.g. [[32]]. Set values to roughly match the expected pixel size of detectable objects.