massachusetts-roads

Massachusetts Roads Benchmark Dataset

High-resolution aerial imagery benchmark for remote sensing road extraction, based on the canonical Massachusetts Roads dataset introduced by Volodymyr Mnih (2013).

Dataset Overview

The Massachusetts Roads Dataset covers diverse urban, suburban, coastal, and forested rural environments throughout the Commonwealth of Massachusetts. - Image Resolution: $1500 \times 1500$ pixels per tile. - Ground Sampling Distance (GSD): $1.0\text{ m/pixel}$. - Channels: 3-channel RGB optical imagery (*_sat.jpg). - Ground Truth: Binary road segmentation masks (*_mask.png), where pixel value 255 (or 1) denotes road corridors rasterized from OpenStreetMap with an approximate line width of 7 pixels, and 0 denotes background.

Benchmark Partition

Standard benchmark partition for road extraction research: | Split | Tiles | Images | Binary Masks | Description | |---|---|---|---|---| | Training (train) | 540 | 540 | 540 | Training set partitioned from the standard 600-image training pool (seed 42) | | Validation (val) | 60 | 60 | 60 | Held-out validation split for model selection & threshold calibration | | Testing (test) | 178 | 178 | 178 | Standard official evaluation benchmark | | Total | 778 | 778 | 778 | Full benchmark corpus |

Directory Structure

massachusetts/
├── metadata.csv                # Complete index (split, tile_id, image_path, mask_path, resolution)
├── train/
│   ├── images/                 # 540 aerial tiles (*_sat.jpg)
│   └── masks/                  # 540 road masks (*_mask.png)
├── val/
│   ├── images/                 # 60 aerial tiles (*_sat.jpg)
│   └── masks/                  # 60 road masks (*_mask.png)
└── test/
    ├── images/                 # 178 aerial tiles (*_sat.jpg)
    └── masks/                  # 178 road masks (*_mask.png)

Quick Start & Usage

1. Fast Download via Python (huggingface_hub)

import zipfile
from huggingface_hub import hf_hub_download

# Download the complete benchmark archive (~530 MB)
zip_file = hf_hub_download(
    repo_id="lammtfkday/massachusetts-roads",
    filename="datasetmassa.zip",
    repo_type="dataset"
)

# Extract into local data folder
with zipfile.ZipFile(zip_file, "r") as zf:
    zf.extractall("./data")

2. Downloading Individual Splits

from huggingface_hub import hf_hub_download

train_zip = hf_hub_download(repo_id="lammtfkday/massachusetts-roads", filename="train.zip", repo_type="dataset")
val_zip   = hf_hub_download(repo_id="lammtfkday/massachusetts-roads", filename="val.zip", repo_type="dataset")
test_zip  = hf_hub_download(repo_id="lammtfkday/massachusetts-roads", filename="test.zip", repo_type="dataset")

3. PyTorch Dataset Loader

import cv2
import torch
from pathlib import Path
from torch.utils.data import Dataset

class MassachusettsRoadDataset(Dataset):
    def __init__(self, root_dir, split="train", transform=None):
        self.img_dir = Path(root_dir) / split / "images"
        self.mask_dir = Path(root_dir) / split / "masks"
        self.images = sorted(list(self.img_dir.glob("*_sat.jpg")))
        self.transform = transform

    def __len__(self):
        return len(self.images)

    def __getitem__(self, idx):
        img_path = self.images[idx]
        mask_path = self.mask_dir / img_path.name.replace("_sat.jpg", "_mask.png")
        image = cv2.cvtColor(cv2.imread(str(img_path)), cv2.COLOR_BGR2RGB)
        mask = cv2.imread(str(mask_path), cv2.IMREAD_GRAYSCALE)
        mask = (mask > 127).astype("float32")
        if self.transform:
            augmented = self.transform(image=image, mask=mask)
            image, mask = augmented["image"], augmented["mask"]
        return torch.tensor(image).permute(2, 0, 1) / 255.0, torch.tensor(mask).unsqueeze(0)

Citations

If you use this dataset in your research, please cite:

@phdthesis{mnih2013aerial,
  title={Machine Learning for Aerial Image Labeling},
  author={Mnih, Volodymyr},
  year={2013},
  school={University of Toronto}
}