Skip to content

Examples

The Quickstart provides an end-to-end example showing how to create an rslearn dataset with aligned Sentinel-2 images and ESA WorldCover land cover labels, how to fine-tune a model on the dataset, and how to apply the model to map land cover in new regions.

Below we list other examples, organized by their focus.

Data-Focused Examples

  • WindowsFromGeojson: create windows based on a GeoJSON file of point features and acquire Sentinel-2 images. Then, train a model to predict the point positions.
  • ProgrammaticWindows: a simple example of creating windows programmatically, in case the dataset add_windows command is insufficient for your use case. This example also shows how to programmatically add raster and vector data into your dataset.
  • NaipSentinel2: create windows based on the timestamp that NAIP is available. Then, acquire NAIP images at each window, along with Sentinel-2 images captured within one month of the NAIP image. This dataset could be used e.g. for super-resolution training.
  • GetImageTimeSeries: a quick example to obtain time series of satellite images for several locations.
  • FindStadiums: a fine-tuning example that shows how to turn point labels into rasters for training a segmentation model, and how to compare several remote sensing foundation models.

Training-Focused Examples

  • FinetuneOlmoEarth: fine-tune OlmoEarth to segment Sentinel-2 images for land cover and crop type mapping. In this example, the labels are taken from the USDA Cropland Data Layer.
  • BitemporalSentinel2: acquire Sentinel-2 images from 2016 and 2024, and train a model to predict which is earlier. This example shows how to specify more complex model architectures (it applies OlmoEarth independently on the two images and then concatenates the feature maps), and also how to add custom augmentations (to randomize the order of the images).

Computing Embeddings

Other Examples