|
|
--- |
|
|
language: en |
|
|
license: mit |
|
|
tags: |
|
|
- environment |
|
|
- computer-vision |
|
|
- vit |
|
|
- climate-change |
|
|
--- |
|
|
|
|
|
# wildfire_smoke_segmentation_vit |
|
|
|
|
|
## Overview |
|
|
This model is a Vision Transformer (ViT) designed for the early detection of wildfires via satellite and aerial imagery. By identifying specific smoke patterns and thermal anomalies, it provides real-time alerts for environmental monitoring agencies. |
|
|
|
|
|
|
|
|
|
|
|
## Model Architecture |
|
|
The model is based on the **ViT-Base** (Patch 16) architecture: |
|
|
- **Patching**: Divides input images into 16x16 patches to capture global spatial dependencies. |
|
|
- **Attention**: Uses multi-head self-attention to distinguish between cloud cover and low-density smoke plumes. |
|
|
- **Pre-training**: Initialized on ImageNet-21k and fine-tuned on the FIRESAT dataset. |
|
|
|
|
|
## Intended Use |
|
|
- **Remote Sensing**: Automated monitoring of vast forested areas via Sentinel-2 or Landsat imagery. |
|
|
- **Early Warning Systems**: Integration into IoT-enabled lookout towers for local fire departments. |
|
|
- **Post-Fire Analysis**: Assessing the spread and intensity of smoke for environmental impact studies. |
|
|
|
|
|
## Limitations |
|
|
- **Atmospheric Conditions**: Heavy cloud cover or fog can lead to false positives. |
|
|
- **Resolution**: Accuracy drops significantly for images where the smoke plume is smaller than 32x32 pixels. |
|
|
- **Time of Day**: Optimized for daytime multi-spectral imagery; night-time performance relies on thermal band availability. |
