NamburiSrinath/satellite-reservoir-estimation-data
satellite-reservoir-estimation — task inputs Agent-visible input bundle for the satellite-reservoir-estimation task in Terminal-Bench-Science. Pulled at container build time, pinned to an immutable commit and verified against per-file SHA-256 sums. This repository deliberately does not identify the site. The task withholds the reservoir's identity from the agent, and this card is reachable from inside the agent container by way of the checksum manifest, so naming the site here… See the full description on the dataset page: https://huggingface.co/datasets/NamburiSrinath/satellite-reservoir-estimation-data.
satellite-reservoir-estimation — task inputs
Agent-visible input bundle for the satellite-reservoir-estimation task in Terminal-Bench-Science. Pulled at container build time, pinned to an immutable commit and verified against per-file SHA-256 sums.
This repository deliberately does not identify the site. The task withholds the reservoir's identity from the agent, and this card is reachable from inside the agent container by way of the checksum manifest, so naming the site here would defeat that. Reviewers: the identity, full provenance and audit path are in the task's own README.md in the benchmark repository, outside the agent image.
Contents
satellite-reservoir-estimation/input/{2009..2018}.tar imagery, one archive per year
satellite-reservoir-estimation/input/dem.tar elevation model
satellite-reservoir-estimation/input/vector.tar reservoir boundary and AOI2,054 files, about 1.0 GB extracted. Shipped as twelve archives rather than loose files because per-request overhead on ~2,000 loose objects costs minutes of build time.
All of it is public domain data from USGS, redistributed unmodified apart from the clipping, mosaicking and attribute stripping described above; imagery pixel values are exactly the publisher's. Band files keep their original USGS names on purpose — the band numbering differs across sensors, and that is part of what the task tests.
Provenance
Both products were retrieved through the Microsoft Planetary Computer STAC API (landsat-c2-l2 and the 3DEP one-metre collection); the elevation tiles are also available directly from prd-tnm.s3.amazonaws.com/StagedProducts/Elevation/1m/. The benchmark task records the exact scene identifiers, the acquisition window and the checksums.
No reference values, expected answers or grading material are in this repository. Those live only in the task's verifier image.
