Source-linked AI summary
Aerial Imagery Pile burn detection using Deep Learning: the FLAME dataset
Alireza Shamsoshoara, Fatemeh Afghah, Abolfazl Razi, Liming Zheng, Peter Z Fulé, Erik Blasch
TL;DR
The paper introduces FLAME, a dataset for pile burns in a Northern Arizona forest, with applications in fire analysis and management. It demonstrates processing algorithms for binary fire classification and fire-mask extraction, showing the dataset’s utility.
Problem
Fire analysis requires improved fire modeling and analysis, while existing global assessments often operate at relatively coarse resolution and have limited repeat imagery.
Method
The paper introduces the FLAME dataset for pile burns in a Northern Arizona forest and applies convolutional neural networks and image-segmentation techniques to fire-labeled data.
Results
The paper provides numerical performance results and demonstrates the utility of FLAME through binary fire classification and fire-mask extraction applications.
Takeaways & Limitations
FLAME supports development of computer tools for fire management and control through annotated fire data and exemplary processing applications.
Takeaways & Limitations
Existing global fire assessments remain constrained by relatively coarse resolution and satellite orbital patterns that limit repeat imagery availability.
Abstract
from arXiv · showhide
Wildfires are one of the costliest and deadliest natural disasters in the US, causing damage to millions of hectares of forest resources and threatening the lives of people and animals. Of particular importance are risks to firefighters and operational forces, which highlights the need for leveraging technology to minimize danger to people and property. FLAME (Fire Luminosity Airborne-based Machine learning Evaluation) offers a dataset of aerial images of fires along with methods for fire detection and segmentation which can help firefighters and researchers to develop optimal fire management strategies. This paper provides a fire image dataset collected by drones during a prescribed burning piled detritus in an Arizona pine forest. The dataset includes video recordings and thermal heatmaps captured by infrared cameras. The captured videos and images are annotated and labeled frame-wise to help researchers easily apply their fire detection and modeling algorithms. The paper also highlights solutions to two machine learning problems: (1) Binary classification of video frames based on the presence [and absence] of fire flames. An Artificial Neural Network (ANN) method is developed that achieved a 76% classification accuracy. (2) Fire detection using segmentation methods to precisely determine fire borders. A deep learning method is designed based on the U-Net up-sampling and down-sampling approach to extract a fire mask from the video frames. Our FLAME method approached a precision of 92% and a recall of 84%. Future research will expand the technique for free burning broadcast fire using thermal images.
1. Introduction: Scope, significance, and problem definition
Wildfires create substantial risks for forests, wildlife, property, and operational personnel, while existing monitoring approaches have important coverage and resolution constraints. The paper introduces FLAME, an aerial fire dataset designed to support early detection, classification, and segmentation research.
- Wildfires have caused severe damage to forests, wildlife habitats, farms, residential areas, and ecosystems.
- Aerial monitoring can provide first responders with more accurate fire-behavior data for fire management.
- Existing fire-monitoring options include personnel, helicopters, fixed-wing aircraft, wireless sensor networks, and satellite imagery.
- Satellite imagery supports global fire assessment but typically has coarse resolution and limited repeat-image availability.
- The FLAME dataset contains drone-acquired fire videos and images from prescribed slash-pile burns in Northern Arizona.
- FLAME includes normal and thermal imagery, supporting frame-based fire classification and pixel-wise fire-mask segmentation.
2. FLAME Dataset: Hardware and Applicable Data
The FLAME dataset documents prescribed slash-pile burning with drone-based aerial imagery collected using multiple cameras and video modalities. It provides normal-spectrum and thermal data, converted video frames, representative fire and no-fire footage, and publicly available dataset resources.
- Dataset collection: The study collected fire imagery during a prescribed slash-pile burn conducted with Flagstaff fire managers in a ponderosa pine forest.The burn occurred on city-owned lands in Arizona under partly cloudy winter conditions.
- Hardware: The dataset was created using different drones, regular cameras, and thermal cameras to capture aerial fire images and videos.Hardware specifications are summarized in Table 1, while the study describes multiple drone and camera types.
- Applicable data: Recorded videos are converted into frames according to their recorded or user-selected FPS to support fire classification and segmentation research.Researchers can use applications of their choice to extract frames at the FPS required by their task.
3. Goals: Suggested Experiments and Methodology
The FLAME dataset supports two related tasks: classifying frames as fire or no-fire and segmenting fire regions pixel by pixel. The paper applies deep neural networks, including Xception for classification and a customized U-Net for fire masking.
- Goals and Problems: The methodology addresses fire-versus-no-fire classification and pixel-wise fire segmentation.Segmentation produces masks of identified fire regions in frames classified as containing fire.
- Fire vs No-Fire Classification: The classification pipeline uses supervised learning on captured camera frames, labeling mixed fire/non-fire frames as fire.Frames without fire are labeled non-fire, and normal-spectrum RGB images are used rather than green or fusion heatmaps.
- Fire vs No-Fire Classification: The binary classifier uses the Xception deep convolutional neural network with a sigmoid output for fire probability.The network uses depth-wise separable convolutions and shortcut connections between convolution blocks.
- Fire vs No-Fire Classification: The classifier is trained with binary cross-entropy and the Adam optimizer, then evaluated on a test dataset.The loss uses ground-truth fire and no-fire labels and predicted fire probabilities.
- Fire Segmentation: The segmentation pipeline treats each pixel as fire or background and trains a customized U-Net using manually generated ground-truth masks.Masks were generated with MATLAB Image Labeler on 2003 frames; the U-Net maps raw images to segmented masks.
- Fire Segmentation: The customized U-Net replaces ReLU with ELU activations and combines up-convolutions with high-resolution features from the contracting path.The input layer is 512 × 512 × 3, with RGB values scaled to floats between 0 and 1.
4. Results: Metrics and guidance on reporting results
The study evaluates separate deep-learning tasks for fire-frame classification and fire-mask segmentation using drone imagery. The reported results include classification accuracy and segmentation precision and recall, with test data collected using different imaging equipment from training data.
- Fire vs No-Fire Classification: The classification training data contained 25,018 fire frames and 14,357 non-fire frames.
- Evaluation Setup: Training and test samples did not overlap because training used a Matrice 200 with a Zenmuse X4S camera, whereas testing used a Phantom drone camera.
- Fire vs No-Fire Classification: 76.23% accuracy was achieved for “Fire vs No-Fire” classification.
- Fire vs No-Fire Classification: Class imbalance produced a higher false-positive rate than false-negative rate in the binary classification task.
- Fire Segmentation: The segmentation evaluation reported precision, recall, AUC, F1-score, sensitivity, specificity, and Mean IOU.
5. Open Challenges regarding the dataset
The dataset supports multiple research directions while leaving open questions about classification accuracy, thermal-image performance, fire-structure segmentation, and broader modeling applications.
- The study identifies two dataset challenges: improving fire classification accuracy and providing more ground-truth masks for fire segmentation.
- GreenHot, WhiteHot, and fused thermal images are available for further fire segmentation and classification investigation.
- Different fire elements, including a white-hot core and exterior regions, could be segmented separately to improve understanding of fire structure.
- Future work includes comparing thermal-image data types to determine which provide better model accuracy.
- Integrating terrain and vegetation-fuel information with imagery could advance image-based fire-spread modeling.
- Further directions include transfer learning, context-based and zero-shot detection, temporal analysis, metric design, user displays, edge efficiency, and occlusion robustness.
6. Conclusion
The paper introduces FLAME as an aerial pile-burn dataset for developing and benchmarking fire-management tools. It demonstrates classification and segmentation applications and encourages broader analyses across imaging modalities and drone-monitoring tasks.
- FLAME is a dataset for pile burns in a Northern Arizona forest, collected with two drones across normal, Fusion, WhiteHot, and GreenHot palettes.
- The dataset supports binary fire classification and fire-mask extraction through two machine-learning applications.
- The demonstrated applications show the utility of FLAME for fire detection and segmentation research.
- FLAME can serve as a benchmark dataset for testing generic image-processing algorithms.
- The paper reports numerical performance results for its image-classification and fire-detection algorithms.
- Future uses include joint analysis of regular and thermal images and segmentation-based task scheduling for drone fleets.