Deep Learning Based Wildfire Detection for Peatland Fires Using Transfer Learning

This paper proposes a transfer learning-based deep learning approach that adapts models pretrained on general wildfire imagery to effectively detect distinct peatland fires using limited labeled data, significantly improving detection accuracy and robustness under challenging conditions like low-contrast smoke and variable illumination.

The Gist

🌋 The Problem: The "Ghost Fire"

Imagine you are a security guard trying to spot a fire.

• Normal Forest Fires are like a bonfire at a campsite. They are loud, bright, and easy to see. The flames are huge, and the smoke is thick and obvious. If you train a robot to spot fires using pictures of these bonfires, it becomes an expert at finding them.
• Peatland Fires (the kind in Malaysia and Indonesia) are different. They are like a slow-burning ember hidden under a blanket. They don't have big, bright flames. Instead, they smolder underground, producing a lot of low-contrast, gray smoke that blends right into the fog or the trees.

The Challenge: The "bonfire expert" robot gets confused by the "ghost fire." It thinks the gray smoke is just fog or clouds and misses the fire. Also, there are very few pictures of these ghost fires to teach the robot, so it can't just learn from scratch.

---

🧠 The Solution: The "Apprentice" Strategy (Transfer Learning)

Instead of trying to teach a robot everything from zero (which takes years and millions of pictures), the researchers used a clever trick called Transfer Learning.

Think of it like hiring a Master Chef who is famous for cooking steaks (general wildfires) and asking them to learn how to cook a specific, tricky regional dish (peatland fires).

1. The Head Start: The robot (the AI model) is first trained on a massive library of photos of normal, bright forest fires. It learns the basics: "What does fire look like? What does smoke look like?"
2. The Specialization: Once the robot is a pro at spotting bonfires, the researchers give it a smaller, specialized cookbook of Malaysian peatland fires. They say, "Okay, you know fire, but now look closely at this specific type of smoldering smoke."
3. The Result: Because the robot already knows what fire generally looks like, it only needs a little bit of extra training to recognize the subtle differences of the peatland fire. It learns much faster and becomes much more accurate than if it had started as a total beginner.

---

🔍 The Tool: The "Super-Spectacles" (WHT-ResNet)

The researchers didn't just use a standard robot; they gave it a pair of specialized glasses called the Walsh-Hadamard Transform (WHT).

• Standard Vision: A normal camera sees pixels (dots of color). If the smoke is gray and the sky is gray, the camera gets confused.
• The WHT Glasses: These glasses don't just look at colors; they look at patterns and structures. Imagine trying to find a specific rhythm in a noisy room. A normal person hears "noise," but someone with perfect pitch hears the specific beat.
• Why it helps: The WHT glasses help the robot ignore the "noise" (like clouds or lighting changes) and focus on the unique "rhythm" of the peatland smoke. It's also lighter and faster, meaning it can run on small, battery-powered cameras in the middle of a remote forest without needing a supercomputer.

---

📸 The Method: The "Puzzle Piece" Approach

Peatland fires often happen in huge, high-resolution images taken by drones or satellites. If you shrink the whole image to make it fit a computer, you might lose the tiny, faint smoke.

The researchers solved this by treating the image like a giant puzzle:

1. They cut the big image into many small squares (patches).
2. They overlapped these squares slightly (like shingles on a roof) so nothing gets lost at the edges.
3. The robot looks at each small square individually to decide, "Is there fire here?"
4. If several neighboring squares say "Yes," the system knows there is a fire.

This way, even a tiny, faint wisp of smoke gets a chance to be seen, rather than getting lost in a giant, blurry picture.

---

🏆 The Results: A Winning Team

When they tested this system, the results were impressive:

• The "Beginner" Robot (trained from scratch) got about 71% of the fires right.
• The "Apprentice" Robot (using Transfer Learning) jumped to 89%.
• The "Apprentice" with "Super-Spectacles" (WHT-ResNet + Transfer Learning) hit 91.6%.

In plain English: By combining the "Master Chef" training (Transfer Learning) with the "Super-Spectacles" (WHT), the system became incredibly good at spotting the tricky, hidden peatland fires. It made fewer mistakes (false alarms) and missed fewer real fires.

💡 Why This Matters

This isn't just about better math; it's about saving forests. Peatland fires release massive amounts of carbon and poison the air for millions of people. By catching these fires early—when they are still just a smoldering ember under the ground—authorities can stop them before they become uncontrollable disasters. This system offers a practical, affordable way to watch over these vulnerable forests 24/7.

Technical Summary

1. Problem Statement

Peatland fires present a unique and critical challenge for wildfire detection systems compared to conventional open-flame forest fires.

• Visual Characteristics: Peat fires are characterized by low-temperature smoldering combustion, minimal or invisible flames, persistent smoke plumes, and subsurface burning.
• Limitations of Current Models: Standard deep learning models trained on general wildfire datasets (dominated by bright, high-contrast flames) fail to detect peat fires effectively because the visual cues (bright fire regions) are often absent or weak.
• Data Scarcity: There is a severe lack of labeled datasets specifically for peatland fires, making it difficult to train deep learning models from scratch without overfitting or poor generalization.
• Operational Challenges: Detection must occur in remote areas with limited computational resources, requiring lightweight models that can handle low-contrast smoke, partial occlusions, and variable illumination.

2. Methodology

The authors propose a framework combining Transfer Learning with a novel Walsh-Hadamard Transform (WHT) based architecture to address data scarcity and visual complexity.

A. Transfer Learning Strategy

• Pre-training: The model is initialized using weights from a deep learning model pre-trained on the Global Wildfire Prevention (GWFP) dataset, which contains diverse open-flame wildfire imagery.
• Fine-tuning: The pre-trained model is then fine-tuned on a curated dataset of Malaysian peatland fire images and videos. This allows the model to adapt its learned feature representations to the specific, subtle characteristics of peat smoke and smoldering fires with minimal additional labeled data.

B. Data Pre-processing (Block Partitioning)

To handle high-resolution images without losing small smoke details (which often occur when resizing images to fit standard network inputs):

• Tiling: Images are divided into overlapping blocks (224×224 pixels).
• Stitching: Four adjacent blocks are stitched to form a sub-image for classification.
• Scoring: The probability score for a specific region is calculated based on the block and its right, bottom, and bottom-right neighbors, ensuring robust detection near edges and avoiding the loss of small smoke signatures.
• Annotation-Free: This approach eliminates the need for manual bounding box labeling, as the system classifies image patches rather than localizing objects via bounding boxes.

C. Architecture: WHT-ResNet50

The paper introduces a hybrid deep neural network that replaces standard convolutional layers with Walsh-Hadamard Transform (WHT) layers within a ResNet50 backbone.

• Mechanism: Instead of standard spatial convolutions (which require complex arithmetic), the network performs convolutions in the transform domain using the orthogonal Hadamard matrix.
• Efficiency: The WHT replaces multiplications with sign-based additions, significantly reducing computational complexity and the number of trainable parameters.
• Hybrid Nature: The network is not fully binary; it is a hybrid architecture combining real arithmetic and binary operations, optimized for feature encoding and lightweight representation learning.

3. Key Contributions

1. Validation of Transfer Learning: Demonstrated that transfer learning is a highly effective strategy for peatland fire detection under limited data conditions, significantly outperforming training from scratch.
2. New Dataset: Presentation of a real-world peatland fire dataset derived from imagery and videos captured in Malaysian peatland regions.
3. Novel Architecture (WHT-ResNet50): Introduction of a WHT-based ResNet50 that leverages the orthogonality of the Hadamard transform to learn better features with fewer parameters compared to standard ResNet.
4. Performance Benchmarking: Showed that the combination of domain-specific pre-training and spectral feature enhancement (WHT) yields superior robustness against false alarms (clouds, fog) and missed detections.

4. Experimental Results

The study evaluated various architectures (EfficientNet-B5, SwinTransformer, ResNet50, HTMA-ResNet50, and the proposed WHT-ResNet50) with and without transfer learning.

• Impact of Transfer Learning:
  • For the baseline ResNet50, accuracy improved from 71.4% (trained from scratch) to 88.9% (with transfer learning).
  • The F1-score improved from 70.1% to 88.2%.
• Impact of WHT Architecture:
  • The WHT-ResNet50 (without transfer learning) already outperformed the standard ResNet50 (77.2% vs. 71.4% accuracy), proving the efficacy of the transform-based feature extraction.
• Best Performance:
  • The WHT-ResNet50 with Transfer Learning achieved the highest performance:
    • Accuracy: 91.6%
    • Precision: 92.9%
    • Recall: 90.17%
    • F1-Score: 91.5%
  • This model also maintained a lower parameter count (20.5M) compared to EfficientNet-B5 (28.3M) and SwinTransformer (27.5M), making it suitable for resource-constrained edge devices.

5. Significance and Conclusion

• Practical Deployment: The proposed solution offers a scalable, real-time monitoring system suitable for remote peatland areas where computational resources are limited.
• Environmental Impact: By improving early detection of smoldering peat fires, the system can mitigate severe environmental consequences, including regional air pollution and long-duration fire suppression costs.
• Generalizability: The approach is not limited to Malaysia; it is applicable to other peat-rich regions (e.g., Indonesia, US) and can be extended to ember detection in beyond-visible imagery.
• Scientific Contribution: The paper establishes that combining transfer learning (to overcome data scarcity) with orthogonal transform-based architectures (to improve feature efficiency) is a superior paradigm for detecting low-contrast, complex wildfire phenomena.