Deep learning model to classify burn scars in multispectral satellite imagery. A brief summary of the item is not available. Add a brief summary about the item.
Deep learning package
by
Item created: Jan 3, 2024 Item updated: Jan 1, 2025 Number of downloads: 3,635
Description
One of the key challenges in monitoring wildfires lies in distinguishing burn scars from non-burnt areas and assessing the extent of damage. This differentiation is crucial to assist emergency responders in their decision-making ability. Satellite imagery enriched with high temporal and spectral information, coupled with advancements in machine learning methods, present an avenue for automated monitoring and management of post-wildfire landscapes on a large scale. The burn scar deep learning model can emerge as an indispensable tool to tackle the task of accurately identifying and mapping the aftermath of wildfires from satellite imagery.
The Prithivi-100M-burn-scar model has been developed by NASA and IBM by finetuning their foundation model for earth observation - Prithvi-100m, using HLS Burn Scar Scenes dataset. Use this model to automate the process of identifying burn scars in multispectral satellite imagery.
Using the model
Follow the guide to
use the model. Before using this model, ensure that the supported deep
learning libraries are installed. For more details, check Deep Learning Libraries Installer for ArcGIS.
Fine-tuning the model
This model can be fine-tuned using the ArcGIS API for Python. Follow the guide to fine-tune this model.
Input
Raster, mosaic dataset, or image service. These must be a 6-band composite raster derived from either Harmonized Landsat 8 (HLSL30) or Harmonized Sentinel 2 (HLSS30). The model can also be used with level-2 products of Sentinel-2 and Landsat-8, yet it performs most effectively with HLSL30 and HLSS30.
The composite raster should contain the following 6 bands: Blue, Green, Red, Narrow NIR, SWIR, SWIR 2.
Band numbers for the above mentioned bands are:
- For HLSS30 and Sentinel-2: Band2, Band3, Band4, Band8A, Band11, Band12
- For HLSL30 and Landsat 8: Band2, Band3, Band4, Band5, Band6, Band7
Output
Classified raster with 2 classes (no burn and burn scar).
Applicable geographies
This model is expected to work well across the globe.
Model architecture
This model packages IBM and NASA's Prithvi-100M-burn-scar and uses a self-supervised encoder developed with a ViT architecture and Masked AutoEncoder (MAE) learning strategy.
Accuracy metrics
This model has an IoU of 0.73 on the burn scar class and 96.00 percent overall accuracy.
Training data
This model finetunes the pretrained Prithvi-100m model to segment the extent of burned area on multispectral satellite images from the HLS Burn Scar Scenes dataset.
Limitations
This model may not give satisfactory results in areas near water bodies.
Sample results
Here are a few results from the model.
Citations
Phillips, C., Roy, S., Ankur, K., & Ramachandran, R. (2023). HLS Foundation Burnscars Dataset.
An in-depth description of the item is not available.
Layers
Tools
Tables
Basemap
Project Contents:
Solution Contents
Contents
Layers
Screenshots
Terms of Use
No special restrictions or limitations on using the item's content have been provided.
Details
Dashboard views: Desktop
Source:
Creating data in:
Published as:
Other Views:
Dependent items in the recycle bin
Applicable: 2d
Size: 1,057.969 MB
ID: 9af7af28dd91473bbc8ad40942e74563
Image Count: 0
Image Properties
Layer Drawing
Using tiles from a cache
Dynamically from data
Share
Owner
Folder
Categories
This item has not been categorized.
Credits (Attribution)
No acknowledgements.IBM, NASA
Comments (0)