| --- |
| license: apache-2.0 |
| tags: |
| - wildfire-detection |
| - computer-vision |
| - cnn |
| - remote-sensing |
| - image-classification |
| - pytorch |
| language: |
| - en |
| --- |
| |
| # PyroSight: Wildfire Detection CNN |
|
|
| **PyroSight** is a convolutional neural network for automated wildfire detection from satellite and aerial imagery. It was developed by Veer Goradia as an independent AI research project. |
|
|
| ## Model Description |
|
|
| Wildfires are among the most destructive natural disasters, and early detection is critical for rapid response. PyroSight uses deep learning to classify remote sensing imagery as wildfire or non-wildfire, enabling automated large-scale monitoring. |
|
|
| **Architecture:** |
| - Custom CNN backbone trained from scratch for wildfire-specific feature extraction |
| - Pretrained baseline comparisons: ResNet50, EfficientNetB0 (frozen + finetuned) |
| - Resolution ablation: 150x150 vs 224x224 input |
| - Saliency map analysis for failure case interpretability |
| - Geographic split analysis to evaluate spatial generalization |
|
|
| **Total accuracy:** 97.62% (deterministic across repeated runs) |
|
|
| ## Performance |
|
|
| | Model | Accuracy | |
| |-------|----------| |
| | PyroSight (custom CNN) | **97.62%** | |
| | ResNet50 (finetuned) | baseline | |
| | EfficientNetB0 (finetuned) | baseline | |
|
|
| Geographic split analysis: 28% of test images within proximity of training regions — model evaluated on strict geographic splits to assess real-world generalization. |
|
|
| ## Training Data |
|
|
| - Satellite and aerial wildfire imagery dataset |
| - Binary classification: wildfire vs. non-wildfire |
| - Geographic split applied to prevent data leakage between train/test sets |
|
|
| ## Intended Use |
|
|
| - Automated wildfire monitoring from satellite imagery feeds |
| - Early warning systems for fire detection |
| - Research on remote sensing and environmental AI |
|
|
| ## How to Use |
|
|
| ```python |
| import torch |
| from tensorflow import keras |
| |
| # Load model |
| model = keras.models.load_model('pyrosight_model') |
| |
| # Predict on an image (224x224 RGB) |
| import numpy as np |
| img = np.random.rand(1, 224, 224, 3) # replace with real satellite image |
| prediction = model.predict(img) |
| print(f"Wildfire probability: {prediction[0][0]:.3f}") |
| ``` |
|
|
| ## Publications |
|
|
| - Submitted to **IEEE Big Data High School Symposium 2026** (Paper ID: SP19206) |
| - Accepted to **Harvard HSRC 2026** |
|
|
| ## Citation |
|
|
| ``` |
| @misc{goradia2026pyrosight, |
| title={PyroSight: Deep Learning for Automated Wildfire Detection from Remote Sensing Imagery}, |
| author={Goradia, Veer}, |
| year={2026} |
| } |
| ``` |
|
|
| ## License |
|
|
| Apache 2.0 — see LICENSE file. |
|
|
| ## Contact |
|
|
| Veer Goradia · vgoradia07@gmail.com |
