Data

The data key of the config.

Input/output data specification for an OlmoEarth model.

Describes the temporal strategy, remote-sensing modalities, and output schema that a model expects. Corresponds to the top-level data key in an olmoearth_config YAML file.

Field	Type	Required
`temporality`	Temporality	yes
`modalities`	Modalities	yes
`output`	Output	yes

Temporality

One of the following, selected by its name field:

`name`	Type
`framed_point_in_time`	FramedPointInTimeTemporality
`period`	PeriodTemporality
`point_in_time`	PointInTimeTemporality

Modalities

Field	Type	Required	Default
`sentinel2_l2a`	Sentinel2L2A	no	`null`
`sentinel1_rtc`	Sentinel1RTC	no	`null`
`landsat_8_9_c2_l1`	Landsat89C2L1	no	`null`

Output

One of the following, selected by its data_type field:

`data_type`	Type
`embeddings`	EmbeddingOutput
`raster`	RasterOutput
`vector`	VectorOutput

PeriodTemporality

Represents the duration of time we're making a prediction for. The observation window equals the prediction period. request.time_of_interest marks the start of the prediction and observation periods.

Field	Type	Required	Default	Description
`name`	`"period"`	yes
`observation`	ObservationStrategy	yes
`allowed_start_month_constraints`	array of Month	no	`null`	Months valid for starting this period.

PointInTimeTemporality

Represents a specific instant we're making a prediction for. Uses a narrow symmetric window centered on request.time_of_interest to find imagery.

Field	Type	Required	Default	Description
`name`	`"point_in_time"`	yes
`observation_window_seconds`	integer (≥ 1)	yes		Total time covered by the symmetric observation window centered on the prediction point.

FramedPointInTimeTemporality

Represents a specific instant we're making a prediction for, with context from before and/or after. Each window contributes its own input layer, so a side may carry several at different resolutions.

Field	Type	Required	Default	Description
`name`	`"framed_point_in_time"`	yes
`lookbehind_observations`	array of ObservationWindow	no	`null`	Observation windows preceding the prediction point.
`lookahead_observations`	array of ObservationWindow	no	`null`	Observation windows following the prediction point.

Sentinel2L2A

Sentinel-2 Level-2A surface reflectance imagery modality.

Field	Type	Required	Default	Description
`sort_by`	Sentinel2L2ASortBy	no	`"cloud_cover"`	The field to sort the data by.
`space_mode`	SpaceMode	yes

Sentinel1RTC

Sentinel-1 radiometrically terrain-corrected (RTC) imagery modality.

Field	Type	Required	Default	Description
`orbit_direction`	Sentinel1RTCOrbitDirection	no	`null`	The direction of the orbit.
`space_mode`	SpaceMode	yes

Landsat89C2L1

Landsat 8/9 Collection 2 Level-1 imagery modality.

Field	Type	Required	Default	Description
`sort_by`	Landsat89C2L1SortBy	no	`"cloud_cover"`	The field to sort the data by.
`space_mode`	SpaceMode	yes

RasterOutput

Field	Type	Required	Default	Description
`data_type`	`"raster"`	yes
`fields`	array of RasterField (min 1 items)	yes		Declared in band order: the first field is band 1, the second band 2, and so on.
`max_pixels_per_dimension`	integer (> 0)	no	`10000`	Maximum pixels per dimension when combining prediction tiles into the output GeoTIFF. Caps output resolution.

VectorOutput

Field	Type	Required	Default	Description
`data_type`	`"vector"`	yes
`fields`	array of VectorField (min 1 items)	yes

EmbeddingOutput

Embedding output.

Embeddings are always power-quantized to int8 (4x compression); the runner adds that value transform automatically. There is nothing to configure.

Field	Type	Required	Default	Description
`data_type`	`"embeddings"`	yes
`max_pixels_per_dimension`	integer (> 0)	no	`10000`	Maximum pixels per dimension when combining embedding tiles into the output GeoTIFF. Caps output resolution.

ObservationStrategy

One of the following, selected by its name field:

`name`	Type
`repeating_interval`	RepeatingIntervalObservation
`sampled`	SampledObservation

Month

Allowed values: "January", "February", "March", "April", "May", "June", "July", "August", "September", "October", "November", "December"

ObservationWindow

One observation window, offset from the prediction point.

Several windows on the same side let a model see one stretch of time at more than one resolution -- coarse periods for the distant past, fine ones for the recent past. They reach the encoder in list order, and their generated layer names follow that order too.

Field	Type	Required	Default	Description
`observation`	ObservationStrategy	yes
`offset_days`	integer (≥ 0)	no	`0`	Gap between this window and the prediction point. For a lookbehind the window ends this many days before the point; for a lookahead it starts this many after.

Sentinel2L2ASortBy

Allowed values: "cloud_cover"

SpaceMode

How to build up or sample imagery.

Allowed values: "mosaic", "contains"

Sentinel1RTCOrbitDirection

Allowed values: "ascending_orbit", "descending_orbit"

Landsat89C2L1SortBy

Allowed values: "cloud_cover"

RasterField

One of the following, selected by its field_type field:

`field_type`	Type
`per_pixel_date`	PerPixelDateField
`per_pixel_regression`	PerPixelRegressionField
`segmentation`	SegmentationField

VectorField

One of the following, selected by its field_type field:

`field_type`	Type
`classification`	ClassificationField
`detection`	DetectionField
`regression`	RegressionField

SampledObservation

Collect discrete snapshots within a window.

Field	Type	Required	Description
`name`	`"sampled"`	yes
`duration_days`	integer (≥ 1)	yes	The observation window duration.
`num_samples`	integer (≥ 1, ≤ 16)	yes	Number of samples to collect.

RepeatingIntervalObservation

Divide window into fixed sub-periods, one observation (mosaic or snapshot) per period.

Field	Type	Required	Default	Description
`name`	`"repeating_interval"`	yes
`num_periods`	integer (≥ 1, ≤ 16)	yes		Number of intervals.
`period_duration_days`	integer (≥ 1)	yes		Duration of each interval.
`recovery_periods`	integer (≥ 0)	no	`0`	Extra periods to search beyond num_periods as fallback for periods lacking satellite coverage. Extends the search window by recovery_periods * period_duration_days. For PeriodTemporality and lookbehind observations, recovery extends backward from the primary window (rslearn prefers primary periods, uses recovery as last resort). Not currently supported on lookahead observations due to rslearn's backward search strategy which would incorrectly prefer recovery periods over primary ones.
`reverse_chronological`	boolean	no	`null`	Deprecated. Use only if you need to preserve old rslearn behavior of loading periods in reverse chronological order to the model.

SegmentationField

A field that represents per-pixel classification

Field	Type	Required	Description
`name`	string (min length 1)	yes	Identifies the field, and matches the model.tasks key that produces it
`field_type`	`"segmentation"`	yes
`allowed_values`	array of ClassValue (min 2 items)	yes

PerPixelRegressionField

Field	Type	Required	Default	Description
`name`	string (min length 1)	yes		Identifies the field, and matches the model.tasks key that produces it
`min_value`	number	yes		Minimum possible value for this field
`max_value`	number	yes		Maximum possible value for this field
`colormap_name`	string	no	`"viridis"`	Name of the rio_tiler colormap to use, see https://cogeotiff.github.io/rio-tiler/colormap/#default-rio-tilers-colormaps
`field_type`	`"per_pixel_regression"`	yes

PerPixelDateField

A per-pixel calendar date, stored as whole days since 1970-01-01 in a uint16 band.

Field	Type	Required	Default	Description
`name`	string (min length 1)	yes		Identifies the field, and matches the model.tasks key that produces it
`field_type`	`"per_pixel_date"`	yes

ClassificationField

A field that represents vector-level classification

Field	Type	Required	Description
`name`	string (min length 1)	yes	Identifies the field, and matches the model.tasks key that produces it
`field_type`	`"classification"`	yes
`allowed_values`	array of ClassValue (min 2 items)	yes

RegressionField

Field	Type	Required	Default	Description
`name`	string (min length 1)	yes		Identifies the field, and matches the model.tasks key that produces it
`min_value`	number	yes		Minimum possible value for this field
`max_value`	number	yes		Maximum possible value for this field
`colormap_name`	string	no	`"viridis"`	Name of the rio_tiler colormap to use, see https://cogeotiff.github.io/rio-tiler/colormap/#default-rio-tilers-colormaps
`field_type`	`"regression"`	yes

DetectionField

A field that represents the classes for detected objects.

Field	Type	Required	Description
`name`	string (min length 1)	yes	Identifies the field, and matches the model.tasks key that produces it
`field_type`	`"detection"`	yes
`detectable_object_classes`	array of ClassValue	yes

ClassValue

A class value for vector classification, with an integer class index.

Field	Type	Required	Description
`value`	integer (≥ 0)	yes	The integer class index (used as the value in GeoJSON label properties)
`label`	string	yes	The human-readable class name
`color`	array or array	yes	The RGB(A) color for this value

Temporality​

Modalities​

Output​

PeriodTemporality​

PointInTimeTemporality​

FramedPointInTimeTemporality​

Sentinel2L2A​

Sentinel1RTC​

Landsat89C2L1​

RasterOutput​

VectorOutput​

EmbeddingOutput​

ObservationStrategy​

Month​

ObservationWindow​

Sentinel2L2ASortBy​

SpaceMode​

Sentinel1RTCOrbitDirection​

Landsat89C2L1SortBy​

RasterField​

VectorField​

SampledObservation​

RepeatingIntervalObservation​

SegmentationField​

PerPixelRegressionField​

PerPixelDateField​

ClassificationField​

RegressionField​

DetectionField​

ClassValue​