NASA Harvest

license

MIT(OSI)

weights

open(data/default_model.pt shipped in the MIT repository and loaded by the package's load_pretrained helper)

data

open(scripts/generate_data.py rebuilds the corpus end to end from Earth Engine, with the tile list, tile statistics and a sample shard shipped in the repository)

code

open(self-supervised train.py, eval.py, configs, and the end-to-end data-generation pipeline)

The weights, the self-supervised pretraining script, and the pipeline that rebuilds the corpus all ship under a plain MIT license with no added attribution or use clause, which is a complete open-source release.

• https://api.github.com/repos/nasaharvest/presto recorded 2026-10-13

  license: MIT; archived: false; fork: false.

• https://api.github.com/repos/nasaharvest/presto/contents recorded 2026-10-13

  Repository root carries train.py, eval.py, mosaiks.py, config, scripts, and the presto package, not inference code alone.

• https://api.github.com/repos/nasaharvest/presto/contents/data recorded 2026-10-13

  Lists default_model.pt at 3.3 MB alongside dw_144_mini_shard_44.tar, dynamic_world_samples.geojson, tile_processing.txt, tile_stats.yaml, and v1_dw_tile_metadata_for_public_release.xlsx.

• https://raw.githubusercontent.com/nasaharvest/presto/main/LICENSE recorded 2026-10-13

  Unmodified MIT License text, Copyright (c) 2024 Presto Authors, with no attribution or field-of-use clause added.

• https://raw.githubusercontent.com/nasaharvest/presto/main/README.md recorded 2026-10-13

  Documents Presto.load_pretrained() reading the state dict from data/default_models.pt, a train.py self-supervised training entry point with a runnable local command, and scripts/generate_data.py generating the training tars from Earth Engine exports, reading the resulting webdataset shards through an authenticated `pipe:gcloud storage cat`. The bucket is not anonymously readable — the object returns 403 and listing returns 401 — so the data dimension is established by the reconstruction scripts rather than by a public download.

The weights ship inside the repository rather than on a model hub, so no download count exists and GitHub stars are used as the last-resort reach signal instead.

• https://api.github.com/repos/nasaharvest/presto recorded 2026-09-10

  stargazers_count: 280; archived: false; fork: false.

Two attributes are stated directly. The encoder-decoder reconstructs masked inputs, which the category does not count as generative, and a variable count of monthly timesteps is one temporal scale rather than several.

• https://raw.githubusercontent.com/nasaharvest/presto/main/README.md recorded 2026-09-10

  Shows construct_finetuning_model adding a linear head for finetuning and a jupyter notebook demonstrating finetuning on different downstream tasks, and specifies the input tensor as the S1, S2, ERA5, and SRTM bands plus Dynamic World classes and latitude and longitude.