Sequoia National Park (SNP) - Hazard Detection Dataset (Wildfire, Extreme Snowfall)

This dataset was collected over a forested mountain region in Sequoia National Park, California, USA, a unique area experiencing both wildfires and extreme out-of-season snowfall throughout the year. Monitoring out-of-season snowfall is critical for nearby communities, as rapid melting can lead to severe flooding. Similarly, wildfire detection is essential for tracking vegetation loss and assessing environmental impact.

Anomaly Labeling

Anomalies in the dataset were identified using two primary criteria:

• Snowfall Anomalies:

  • Images captured when snow depth exceeded the maximum or fell below the minimum recorded during the training period were labeled as anomalies.
  • Snow depth data was obtained from the Leavitt Meadows ground station via the National Water and Climate Center (NWCC).

• Wildfire Anomalies:

  • Visual inspection of RGB and infrared images was used to identify wildfire-affected areas.
  • Additionally, images with an average Normalized Burn Ratio (NBR) below the training minimum were labeled as anomalies, indicating significant vegetation loss (i.e., burned regions).

Data Collection & Processing

The dataset consists of imagery collected through the Sentinel-2 satellite constellation and extracted using Sentinel Hub's Python API. The dataset spans different geographical regions of interest, each featuring various types of hazards.

• Processing Level:

  • All images use Level-2A processing, meaning pixel values are radiometrically corrected to provide bottom-of-atmosphere surface reflectance.

• Spectral Bands Used:

  • Of the 13 available Sentinel-2 bands, only the 10 bands at 10m or 20m resolution were used.
  • The three 60m resolution bands (primarily used for atmospheric correction) were excluded.

• Spatial Resolution:

  • The dataset was downsampled to 30m per pixel resolution, balancing large-scale hazard detection with adequate regional coverage.
  • Each image has a spatial resolution of 1024 × 1024 pixels.

Dataset Composition

• Training Set:

  • Contains 186 images (each stored as a .npy file with a corresponding .png file).
  • Captured between January 2017 and July 2020.

• Test Set:

  • Contains 250 images.
  • Captured between August 2020 and August 2023.

• Cloud Filtering:

  • Images with more than 30% cloud coverage, as defined by Sentinel-Hub, were removed.
  • Additional images with visually high cloud coverage were manually filtered out.