Akash-LISS4-CRPaired cloudy and clear LISS-IV tiles over North-East India, with Sentinel-1 radar, terrain and exact ground truth
Start with a 3.4 GB benchmark, train on 244,473 curated tiles, scale to 2.55 million, or take the whole 4.9 TB archive.
from datasets import load_dataset ds = load_dataset("khamirdesai/Akash-LISS4-CR", "benchmark", split="test")
DeliveredReconstructed
Ridges east of Gangtok, 27 Jan 2025, ≈20 km at 20 m, false colour (R = NIR). Map mode, which also reads the 4 Dec 2023 clear pass.
Pick the size you need
Every tier is the same data at a different scope, and each tile carries its location, both acquisition dates, archive shard and position, so you can always trace it back.
Bars show size on one log scale. Every tier is complete; row counts are checked shard by shard against the index.
What a tile carries
Twelve benchmark tiles, straight from the published Parquet, with the release model's output beside the exact target. They are picked for visible structure from the highest margins over the floor in each regime, so they show the best of the method. The median tile scores 38.2 dB under thin cloud, 35.8 dB under ordinary cloud and 30.6 dB under opaque cloud.
Columns and arrays
Arrays are stored as .npy bytes, so values and dtypes are exact. Decode with numpy.load(io.BytesIO(row[col])).
| Column | benchmark · realcloud | curated · pool | Meaning |
|---|---|---|---|
| cloudy | float16 [3,256,256], DN/1023 | uint16 [3,256,256], 10-bit DN | Cloudy LISS-IV tile, bands B2 B3 B4 (green, red, NIR) |
| target · clear | float16 [3,256,256] | uint16 [3,256,256] | Exact ground truth (benchmark) or the clear acquisition of the same ground |
| cloud_opacity · alpha | float16 [256,256] | float16 [256,256] | Real cloud opacity α in [0, 1] |
| radar | float16 [2,256,256] | float16 [2,64,64], 20 m | Sentinel-1 VV, VH γ⁰ (dB in [−25, 0] scaled to [0, 1]) |
| terrain | float16 [2,256,256] | float16 [2,64,64], 20 m | GLO-30 elevation and hillshade. Curated and pool store elevation with 1.5 headroom: divide by 1.5 and clip to [0, 1] |
| pair_id, pair_name, location | string | string | Pair key (location | clear date | cloudy date) and path/row/sub-scene |
| clear_date, cloudy_date | string | string | Acquisition dates, ISO 8601 |
| cloud_class, regime | string | cloud_class | heavy / moderate / light, from pairing; regime thin / ordinary / opaque (benchmark) |
| archive_shard, shard_tile, y, x | string, int | string, int | Where the tile lives in the archive and its position in the scene |
| split, coherent, weight, harmonisation_gain_bias | — | curated only | Training split, coherence flag, sampling weight, per-band gain and bias [3,2] |
Copy, paste, run
Train and test share no acquisition
Clear acquisitions recur across pairs, so a split by pair leaks: 80.4% of validation tiles would have a ground truth that is also a training target. Every split here is acquisition-disjoint from the 46 test pairs.
| Set | Tiles | Pairs | Locations | Note |
|---|---|---|---|---|
| benchmark | 3,008 | 46 | — | Exact targets by cloud transplantation; 576 tiles from 9 pairs at locations absent from training |
| realcloud | 259 | 9 | — | Genuinely cloudy tiles with a harmonised clear partner; score without a reference |
| curated train | 166,319 | 270 | 53 | Heavy 107,193 · moderate 73,819 · light 63,461 across all three |
| curated validation | 62,799 | 18 tiles also in train, counted there | ||
| curated test_pair | 15,355 | Tiles of the test pairs (not the benchmark itself) | ||
| pool | 2,546,219 | 413 | 52 | Every distinct tile sharing no acquisition with the test pairs |
Licence
LISS-IV imagery and everything derived from it: Source: NRSC, ISRO/DOS (www.nrsc.gov.in), non-commercial use under the NRSC terms. Radar and terrain contain modified Copernicus Sentinel data and the Copernicus DEM GLO-30.
Known limits
- One region, North-East India, 2012–2026.
- 9.0% of benchmark tiles have no radar; pairs before October 2014 use later Sentinel-1 passes.
- Cloud classes come from a quicklook at pairing time, not a calibrated cloud product.
Every model trained on this data
From the release cascade down to every intermediate run, on LISS-IV and on SEN12MS-CR: 168 distinct checkpoints, with weights for 90 of them in safetensors and one loader.
Release cascade
34.95 dBFusionNet R44 + DehazeNet2, the only method above the 33.07 dB floor. release/
CloudBridge
32.91 dBDiffusion bridge on FusionNet. Stochastic, so not in the cascade. weights
FusionNet, equal budget
31.20 dBBest of the five LISS-IV fine-tunes; still below the floor. weights
Scored in the paper
| Model | Trained on | LISS-IV PSNR | LISS-IV zero-shot | SEN12MS-CR PSNR | Weights |
|---|
LISS-IV: 3,008 benchmark tiles, floor 33.07 dB. SEN12MS-CR: n = 825, every model trained by us. DSen2-CR, GLF-CR, UnCRtainTS and DiffCR are our re-implementations.
All 168 checkpoints
| Checkpoint | Architecture | Parameters | Run best | Download |
|---|
BibTeX
@inproceedings{desai2026akashliss4cr,
title = {Akash-{LISS4}-{CR}: A Multimodal Benchmark and Reconstruction Cascade for Cloud Removal
in High-Resolution {LISS-IV} Imagery},
author = {Desai, Khamir and Rathod, Daksh and Tejot, Harshil and Chouhan, Avinash and
Lyngdoh, Rosly B and Kumar, Pankaj and Mehta, Mayuri},
booktitle = {Proceedings of the 17th Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP)},
year = {2026}, address = {Kolkata, India}, publisher = {ACM}
}Khamir Desai¹, Daksh Rathod², Harshil Tejot², Avinash Chouhan³, Rosly B Lyngdoh³, Pankaj Kumar², Mayuri Mehta¹ · ¹ SCET Surat · ² Dhirubhai Ambani University · ³ NESAC, Umiam. We thank NRSC, ISRO, for the LISS-IV data distributed through Bhoonidhi, and ESA and the Copernicus programme for Sentinel-1 and GLO-30.