DatasetEnum
Available datasets for automated ingestion and mosaic creation. Seasonal composites use latitude-based heuristics to determine day-of-year (DOY) ranges to select valid observations for compositing.Values
Sentinel-2 median composite dataset for harvest season.
Sentinel-2 median composite dataset for planting season.
InferenceActorEnum
Available inference actors that run models. Each actor handles a particular task, typically accepting a specific input signature and returning a specific output signature.Values
PyTorch-based regression inference actor for continuous value prediction tasks. Compatible with models that accept a single tensor input and return a tensor of continuous values.
PyTorch-based semantic segmentation inference actor for pixel-level classification tasks. Compatible with models that accept a single tensor input and return a tensor of logits that can be softmaxed to produce per-class probabilities.
InferenceConfig
Configuration for model inference in RasterFlow.Attributes
Path to the model file used for inference. Must be a Pytorch 2 Model Archive and have the extension .pt2. See https://docs.pytorch.org/docs/stable/export/pt2_archive.html for more information.
The inference actor type to use for running the model.
Size of the patches to be used during inference in pixel units.
Size of the clip to be used during inference in pixel units.
Device to run the model on (e.g., “cuda”, “cpu”).
List of feature names to be used for the model inference. This must match the underlying band labels of the data source.
List of label names to be used for the model inference.
Maximum batch size to use during inference.
The mode to use for merging overlapping patches.
MergeModeEnum
Available patch overlap merge modes for inference.Values
Merge overlapping patches by clipping to patch boundaries without blending. Produces sharper boundaries between patches.
Merge overlapping patches using weighted averaging based on distance from patch edges. Produces smoother transitions between patches.
ModelRecipes
Available model recipes for inference workflows.Values
Recipe using the FTW model for inference. This model works on an annual basis using Sentinel-2 to derive planting and harvest dates for agricultural fields based on latitude using a greedy search to find up to 10 valid observations in the planting and harvest windows based on the underlying QF band and finally takes the median across those observations to derive the final planting and harvest pixel values. The input dataset is natively in a UTM projection.
Recipe using the Meta CHM v1 model for inference. Please note that this is a best-effort recipe given underlying NAIP data quality and availability. For example, we filter to NAIP at 60 cm resolution, but it is up to the user to specify the start/end dates and aoi that will yield valid NAIP data. Refer to the following link for available NAIP data coverage: https://www.arcgis.com/sharing/rest/content/items/7198c7b7e9854dcaaf36f42378dc8bc4/data The input dataset is natively in a UTM projection.
Recipe using the Tile2Net model for inference. Please note that this is a best-effort recipe given underlying NAIP data quality and availability. For example, we filter to NAIP at 30cm resolution, but it is up to the user to specify an AOI and start/end dates that will yield valid NAIP data. refer to the following link for available 30cm NAIP data coverage: https://www.arcgis.com/sharing/rest/content/items/7198c7b7e9854dcaaf36f42378dc8bc4/data The input dataset is natively in a UTM projection.
ModelRegistryConfig
Configuration for model registry settings.Attributes
Path to the model file used for inference. Must be a Pytorch 2 Model Archive and have the extension .pt2. See https://docs.pytorch.org/docs/stable/export/pt2_archive.html for more information.
The inference actor to use for inference.
Size of patches in pixels. Defaults to None.
Size of clips in pixels. Defaults to None.
Device to run on (e.g., “cuda”, “cpu”). Defaults to None. Note: While “cpu” is supported, we only support GPU runtimes at this time.
List of feature names. Defaults to None.
List of label names. Defaults to None.
Maximum batch size for inference. Defaults to None.
Mode for merging overlapping patches. Defaults to None.
ModelRegistryEnum
Available model registries for model storage and retrieval.Values
HuggingFace model registry for accessing publicly available models.
MLM (Machine Learning Model) registry for internal model storage.
ResamplingMethod
Enumeration of available resampling methods for raster data. See: https://gdal.org/en/latest/drivers/raster/gti.html#tile-index-requirementsValues
Average resampling. Computes the average of all non-NODATA contributing pixels. Good for downsampling.
Cubic convolution resampling. Produces smoother results than nearest neighbor. Good for continuous data.
Cubic spline resampling. Smoother than cubic, higher computational cost.
Gaussian kernel resampling. Produces smooth results with gaussian weighting.
Lanczos windowed sinc resampling. High quality, slower performance. Excellent for preserving sharp features.
Nearest neighbor resampling. Fastest method, preserves original values. Best for categorical data.
Root mean square resampling. Computes RMS of all non-NODATA contributing pixels.
RuntimeEnum
Available runtime configurations for compute resources.Values
SemSegRasterioConfig
Configuration for semantic segmentation vectorization using rasterio.Attributes
Whether to compute and include statistics about the segmentation output. Defaults to True.
Whether to apply medial axis skeletonization to the segmentation mask. Defaults to False.
VectorizeMethodEnum
Available vectorization methods for converting raster outputs to vector formats.Values
Vectorization method using rasterio for semantic segmentation outputs. Converts classified raster pixels to polygon geometries.