sattwik21/mmls-v2

MMLS v2: Multimodal Martian Landslide Dataset

This repository hosts a Hugging Face compatible, Parquet-formatted version of the MMLS v2 dataset. It contains 128x128 multi-channel images for Martian landslide segmentation, featuring 7 data channels and 1 segmentation mask.

Disclaimer: I am not the original creator of this dataset. I am hosting this pre-processed version to allow researchers to easily stream the multi-channel arrays directly into PyTorch training pipelines without dealing with manual TIFF decoding. All credit for data collection and the original paper belongs to the authors of MMLSv2. Please see the Citation section below.

Dataset Structure

Each sample in the dataset is a dictionary containing pre-formatted, multi-dimensional arrays. To maintain 3D spatial consistency across all modalities, single-channel arrays include an explicit channel dimension of 1.

• rgb: (3, 128, 128) — Float32
• dem: (1, 128, 128) — Float32
• thermal_inertial: (1, 128, 128) — Float32
• grayscale: (1, 128, 128) — Float32
• label: (128, 128) — Float32 (Segmentation Mask)

Quickstart (PyTorch)

You can load this dataset directly into PyTorch tensors using the standard Hugging Face datasets library. No manual data conversion is required.

from datasets import load_dataset

# 1. Load a specific split from the Hub (train, val, or test)
dataset = load_dataset("sattwik21/mmls-v2", split="train")

# 2. Automatically format all arrays as native PyTorch Tensors
dataset = dataset.with_format("torch")

# 3. Pull a sample to verify
sample = dataset[0]

# 4. Verify tensor dimensions
print("RGB shape:", sample["rgb"].shape)                 # Expected: torch.Size([3, 128, 128])
print("DEM shape:", sample["dem"].shape)                 # Expected: torch.Size([1, 128, 128])
print("Thermal shape:", sample["thermal_inertial"].shape)# Expected: torch.Size([1, 128, 128])
print("Grayscale shape:", sample["grayscale"].shape)     # Expected: torch.Size([1, 128, 128])
print("Label shape:", sample["label"].shape)             # Expected: torch.Size([128, 128])

References

@inproceedings{paheding2026mmlsv2,
    title={Mmlsv2: A multimodal dataset for martian landslide detection in remote sensing imagery},
    author={Paheding, S. and Reyes-Angulo, A. A. and Ramos, L. T. and Sappa, A. D. and KS, S. K. and Oommen, T.},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    pages={10329--10338},
    year={2026}
  }