# Akash-LISS4-CR

> Akash-LISS4-CR is the first public paired cloud-removal dataset and benchmark for the Indian Resourcesat-2/2A LISS-IV sensor (three VNIR bands, 5.8 m), over North-East India, 2012-2026. Every tile has co-registered Sentinel-1 VV/VH radar and Copernicus GLO-30 terrain. The benchmark has 3,008 test tiles with exact, same-date ground truth made by transplanting real LISS-IV cloud onto clear tiles of the same ground. Accepted at ICVGIP 2026 (ACM).

Use it for: any LISS-IV research (2,790,692 co-registered 256 px tiles); benchmarking optical-SAR-DEM multimodal architectures on a fixed protocol with an acquisition-disjoint split; and pretraining (the 2,546,219-tile `pool` shares no acquisition with the test pairs). Licence: imagery non-commercial under the NRSC terms ("Source: NRSC, ISRO/DOS"); code MIT.

Load it: `load_dataset("khamirdesai/Akash-LISS4-CR", "benchmark", split="test")`. Configs: benchmark (3,008), realcloud (259), curated (244,473), pool (2,546,219).

Cite both the dataset and the paper (BibTeX on the project page). Dataset: Desai, K., Rathod, D., Tejot, H., Chouhan, A., Lyngdoh, R. B. and Kumar, P. (2026). Akash-LISS4-CR: A Multimodal LISS-IV Cloud-Removal Dataset and Benchmark (Version 1.0) [Data set]. Hugging Face. DOI: https://doi.org/10.57967/hf/11073. Paper: Desai, K., Rathod, D., Tejot, H., Chouhan, A., Lyngdoh, R. B., Kumar, P. and Mehta, M. Akash-LISS4-CR: A Multimodal Benchmark and Reconstruction Cascade for Cloud Removal in High-Resolution LISS-IV Imagery. ICVGIP 2026, ACM (to appear).

## Data
- [Dataset on Hugging Face](https://huggingface.co/datasets/khamirdesai/Akash-LISS4-CR): configs, fields, splits, scorer, per-tile scores
- [Data portal](https://huggingface.co/spaces/khamirdesai/Akash-LISS4-CR-Data): download tiers, tile explorer, fields, verification, FAQ
- [Full archive, 4.90 TB](https://huggingface.co/datasets/khamirdesai/Akash-LISS4-CR-archive): raw scenes, every tile shard, radar and terrain stores, every run

## Models
- [Models](https://huggingface.co/khamirdesai/Akash-LISS4-CR-models): 168 checkpoints catalogued, 95 in safetensors; flagship release cascade 34.95 dB PSNR on the benchmark, the only method above the 33.07 dB mean-colour floor

## Paper and code
- [Project page](https://huggingface.co/spaces/khamirdesai/Akash-LISS4-CR)
- [Code](https://github.com/Eartherai/Akash-LISS4-CR/tree/main): benchmark harness, verification scripts, paper source