Source-linked AI summary
Active Fire Detection in Landsat-8 Imagery: a Large-Scale Dataset and a Deep-Learning Study
Gabriel Henrique de Almeida Pereira, André Minoro Fusioka, Bogdan Tomoyuki Nassu, Rodrigo Minetto
TL;DR
Active fire detection is important for environmental monitoring, but deep-learning research lacks large-scale datasets and evaluation architectures. The paper constructs public Landsat-8 datasets and trains CNNs to approximate handcrafted algorithms, including reduced-band and combined-output models. Combined CNN outputs outperform the original handcrafted conditions, while the resulting complex relationships are harder to understand.
Problem
Active fire detection supports environmental monitoring, but deep-learning research lacks large-scale datasets and architectures for evaluation.
Method
The paper constructs public Landsat-8 datasets and trains CNN architectures to approximate three handcrafted algorithms, including reduced-band models and combined outputs.
Results
Combined CNN outputs achieve better performance than the original handcrafted conditions, and CNNs approximate them using only three channels.
Takeaways & Limitations
The datasets, source code, and trained models provide public resources for benchmarking active-fire detection architectures and approaches.
Takeaways & Limitations
Handcrafted rules are sensitive to small threshold changes, while CNNs encode more complex but harder-to-understand relationships; robust CNN architecture design is not the paper’s main focus.
Abstract
from arXiv · showhide
Active fire detection in satellite imagery is of critical importance to the management of environmental conservation policies, supporting decision-making and law enforcement. This is a well established field, with many techniques being proposed over the years, usually based on pixel or region-level comparisons involving sensor-specific thresholds and neighborhood statistics. In this paper, we address the problem of active fire detection using deep learning techniques. In recent years, deep learning techniques have been enjoying an enormous success in many fields, but their use for active fire detection is relatively new, with open questions and demand for datasets and architectures for evaluation. This paper addresses these issues by introducing a new large-scale dataset for active fire detection, with over 150,000 image patches (more than 200 GB of data) extracted from Landsat-8 images captured around the world in August and September 2020, containing wildfires in several locations. The dataset was split in two parts, and contains 10-band spectral images with associated outputs, produced by three well known handcrafted algorithms for active fire detection in the first part, and manually annotated masks in the second part. We also present a study on how different convolutional neural network architectures can be used to approximate these handcrafted algorithms, and how models trained on automatically segmented patches can be combined to achieve better performance than the original algorithms - with the best combination having 87.2% precision and 92.4% recall on our manually annotated dataset. The proposed dataset, source codes and trained models are available on Github (https://github.com/pereira-gha/activefire), creating opportunities for further advances in the field
1 Introduction
Active fire detection supports environmental monitoring, conservation policy, and related tasks, but deep-learning research remains constrained by limited large-scale datasets and evaluation architectures. This paper addresses those gaps with public datasets, CNN experiments, algorithm comparisons, and released code.
- Motivation: Satellite-based active fire detection supports conservation policy, decision-making, and monitoring of deforestation, biomass burning, climate, damage, prevention, and prediction.Fires may be used to clear land, linking active-fire presence to accelerated deforestation in important biomes.
- Existing approaches: Active-fire methods typically determine whether multispectral pixels represent fire using sensor-specific conditions, thresholds, and neighborhood statistics.Different satellite altitudes, sensors, and wavelengths require methods designed for particular systems, including Landsat-8.
- Research gap: Deep learning for active fire recognition remains relatively new and lacks large-scale datasets and architectures for evaluation.The paper identifies this as a central gap motivating its contributions.
- Contributions: The paper introduces a public dataset of 146,214 Landsat-8 patches with 10-band imagery and outputs from three established handcrafted fire-detection algorithms.A secondary public dataset contains 9,044 patches from 13 September 2020 images with manually annotated fire pixels.
- Contributions: The study evaluates CNNs as approximations of handcrafted algorithms, reduced-band models, and combinations of multiple outputs against manual annotations.The authors also release source code for the deep-learning and handcrafted algorithms.
- Paper organization: The paper is organized around dataset generation, CNN architectures, experiments, discussion, and conclusions with future work.Datasets are described in Section 2, architectures in Section 3, experiments in Section 4, and conclusions in Section 5.
2 Materials
The materials comprise globally distributed Landsat-8 imagery from August 2020, processed into patches and segmentation masks using three established rule-based algorithms. The resulting public dataset preserves diverse geographic conditions while exposing algorithmic disagreement for benchmarking.
- Temporal coverage: August and September were selected because fire seasons can occur concurrently in both hemispheres, with these months identified as especially critical.The selection accounts for differing seasonal and climatic behavior across hemispheres.
- Mask generation: Because manually annotating the full imagery volume was infeasible, three established rule sets generated automatic segmentation masks for training data.The rules were developed by Schroeder et al., Murphy et al., and Kumar and Roy.
- Detection algorithms: Schroeder’s conditions use seven channels, classify pixels through conjunctive rules, and relax detection using statistics from a 61 × 61 neighborhood.Neighborhood means and standard deviations exclude fire and water pixels, which are identified by additional conditions.
- Detection algorithms: Murphy’s non-contextual conditions use channels c5, c6, and c7, while Kumar and Roy’s method uses channels c2–c7 with progressively larger contextual neighborhoods.Kumar and Roy test neighborhoods from 5 × 5 through 61 × 61 until sufficient unclassified pixels are available.
- Dataset characteristics: The three algorithms generally agree on fire presence but differ in pixel-level segmentation, while the dataset spans diverse global environments and challenging scenarios.Covered settings include deserts, rain forests, agricultural land, water, snow, clouds, cities, and mountains.
3 Methods
The methods evaluate U-Net-based CNNs for approximating conventional active-fire detection outputs using Landsat-8 imagery with either 10 or 3 input bands. The study compares architecture variants, binary-output thresholding, and multiple target-mask construction strategies.
- Data and masks: The methods use image patches and segmentation masks from the proposed dataset, displaying channels c7, c6, and c2 while retaining 10 bands in the original patches.The dataset examples show that the three algorithms can agree on fire presence while differing in pixelwise segmentation.
- CNN architectures: U-Net combines downsampling for feature and context extraction with upsampling and skip connections for precise pixel-level localization.The tested implementation also adds batch normalization and dropout layers.
- CNN architectures: Three U-Net variations were evaluated, including 10-channel and 3-channel inputs using Landsat-8 bands c7, c6, and c2.The reduced-band configuration tests whether comparable results can be obtained with lower bandwidth, memory, and storage requirements.
- Output masks: Each CNN produces a 256 × 256 binary mask, labeling pixels as fire when its output exceeds the empirically selected threshold of 0.25.Values above 0.25 are set to 1 and other values to 0.
- Training targets: The models were trained and tested across 15 scenarios formed from three handcrafted condition sets, their intersection, and a best-of-three voting mask.The voting mask labels a pixel as fire when at least two condition sets agree.
4 Results and Discussion
The experiments show that CNNs can approximate handcrafted active-fire segmentation, including with reduced three-channel architectures, and that combining CNN outputs can improve performance against manual annotations. Persistent errors remain in some settings, while learned models trade interpretability for more complex decision rules.
- All architectures reproduced the handcrafted algorithms reasonably well, with three-channel models often performing similarly to the larger ten-channel U-Net.The results suggest that channels c7, c6, and c2 contain most information needed for active-fire detection.
- The intersection of three network outputs increased precision but sharply reduced recall, whereas voting reduced algorithm-specific behaviors while retaining stronger recall.The intersection keeps only pixels all networks classify as fire; voting addresses behaviors such as thicker clusters or holes produced by individual condition sets.
- Compared with handcrafted conditions, CNNs generally improved recall over the Schroeder and Kumar–Roy methods without a corresponding precision loss, while maintaining a precision–recall trade-off against Murphy’s conditions.
- The best overall performance came from voting across three networks, with the light three-channel U-Net most similar to manual segmentations and no significant difference from the larger three-channel U-Net.These networks were trained only on automatically segmented samples, while manually annotated patches were used for testing.
- Some errors persisted across approaches, including false detections around urban areas and threshold-sensitive mistakes caused by reflectance values near handcrafted-rule boundaries.Temporal analysis was suggested for persistent urban-area errors, while machine-learned rules can encode more precise but less interpretable relationships.
5 Conclusions
The paper introduces large public Landsat-8 datasets for active fire detection and evaluates CNNs as approximations of handcrafted algorithms. CNN ensembles can outperform individual handcrafted condition sets, while more complex learned rules are harder for humans to understand.
- 146,214 Landsat-8 image patches and 9,044 manually annotated patches provide public resources for training and evaluating active fire detection methods.The first dataset includes outputs from three established handcrafted algorithms; the second contains human-produced annotations.
- Combining multiple CNN outputs achieves better performance than individual handcrafted condition sets.The improvement is reported as a consequence of combining model outputs rather than relying on one condition set.
- CNNs approximate handcrafted active fire algorithms even with reduced architectures using only three sensor channels.The study reports that this reduced-channel approach can encode the relevant detection behavior.
- Deep learning can encode more complex rules than fixed handcrafted thresholds, but those learned relations are harder to understand and describe.The paper frames this as a trade-off between improved detection behavior and interpretability.
- Cross-satellite performance remains an open question for future work, including whether models trained on one satellite transfer effectively to others.The paper also proposes combining results from satellites with different responses, resolutions, and revisit cycles.