💻 computer science

U-Net with Hadamard Transform and DCT Latent Spaces for Next-day Wildfire Spread Prediction

This paper introduces TD-FusionUNet, a lightweight deep learning model that integrates trainable Hadamard and Discrete Cosine Transform layers with custom preprocessing techniques to achieve efficient and accurate next-day wildfire spread prediction using multimodal satellite data.

Original authors: Yingyi Luo, Shuaiang Rong, Adam Watts, Ahmet Enis Cetin

Published 2026-02-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yingyi Luo, Shuaiang Rong, Adam Watts, Ahmet Enis Cetin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict where a wildfire will spread tomorrow. It's like trying to guess the path of a chaotic, dancing flame based on the wind, the dry grass, and the shape of the hills. Traditional methods are like trying to solve this puzzle by manually calculating every single step of the fire's physics—slow, heavy, and often missing the big picture.

This paper introduces a new, lightweight "super-forecaster" called TD-FusionUNet. Think of it as a smart, efficient detective that doesn't just look at the fire; it listens to the "music" of the landscape to predict the next move.

Here is how it works, broken down into simple concepts:

1. The Problem: Too Much Noise, Not Enough Data

Wildfire data is tricky. Most of the map is just "safe ground" (unburned), and only tiny, scattered spots are "danger zones" (burned). It's like trying to find a few specific red marbles in a giant bucket of blue ones. Also, standard AI models that try to solve this are often like over-eating elephants: they are huge, slow, and need massive computers to run, which isn't helpful if you need an answer right now in a remote forest.

2. The Solution: The "Frequency" Detective

The authors built a new AI model that uses a clever trick. Instead of looking at the image pixel-by-pixel (like looking at a painting dot-by-dot), they ask the AI to look at the patterns and frequencies.

The Analogy: Imagine listening to a song. A standard AI tries to memorize every single sound wave. This new AI, however, uses a Hadamard Transform and a DCT (Discrete Cosine Transform) to break the song down into its musical notes (frequencies).
Why this helps: Just as a song is defined by its melody and rhythm rather than just the noise, a wildfire spread is defined by specific patterns (like how fire moves up a hill or follows the wind). By focusing on these "musical notes" (the essential frequencies), the AI can ignore the static noise and learn the rules of fire spread much faster and with less memory.

3. The "Two-Brain" Architecture

The model is called TD-FusionUNet (Transform Domain Fusion). Think of it as a detective with two specialized brains working together:

Brain A (Hadamard): Good at spotting sharp, blocky patterns and sudden changes.
Brain B (DCT): Good at spotting smooth, gradual curves and waves.

They both look at the same data but through different "lenses." Then, they combine their findings. This is like having one expert who is great at spotting sharp corners and another who is great at spotting curves; together, they see the whole picture perfectly.

4. Cleaning Up the Messy Data

Since fire data is so sparse (mostly empty space), the researchers invented a special "pre-processing" technique to help the AI learn better:

Random Margin Cropping: Imagine taking a photo of a fire and randomly zooming in and out or shifting the frame slightly. This forces the AI to learn the shape of the fire, not just memorize its exact location.
Gaussian Smoothing: They take the sharp, jagged edges of the fire map and gently blur them, like softening a rough sketch. This helps the AI understand that fire doesn't jump in perfect squares; it spreads in soft, organic waves. This creates "synthetic" examples so the AI doesn't get confused by missing data.

5. The Results: Small but Mighty

The best part? This super-smart detective is incredibly small.

The Old Way: The previous best models were like heavy trucks (14 million "parameters" or brain cells). They were accurate but slow and required powerful servers.
The New Way: This new model is a compact sports car (only 370,000 parameters). It is about 40 times smaller than the competition but actually more accurate.

Why Does This Matter?

Because this model is so small and efficient, it can run on edge devices. This means you could potentially run this prediction tool on a drone, a satellite, or a handheld device in a remote forest without needing a supercomputer.

In a nutshell: The authors built a tiny, efficient AI that listens to the "music" of wildfire patterns instead of staring at the pixels. By using two different mathematical "lenses" and cleaning up the messy data, it predicts where fires will go tomorrow with high accuracy, using a fraction of the computing power required by older methods. This could save lives by getting fire crews to the right place, faster.

1. Problem Statement

Wildfires are increasingly frequent and severe due to climate change, causing massive economic losses and casualties. Effective disaster response relies on accurate next-day wildfire spread prediction to allocate resources optimally.

Challenges: Traditional physics-based simulators struggle with complex, nonlinear environmental interactions. Existing deep learning approaches (e.g., standard CNNs, UNets) often suffer from:
- High computational costs and large parameter counts, making them unsuitable for edge devices or resource-limited environments.
- Data scarcity and imbalance, particularly the sparsity of "fire" pixels compared to "unburned" areas in satellite imagery.
- Inefficiency in capturing essential spatial-frequency patterns required for accurate segmentation.

2. Methodology

The authors propose TD-FusionUNet (Transform Domain Fusion UNet), a lightweight, multi-modal deep learning architecture designed to balance accuracy and efficiency.

A. Core Architecture: Transform Domain Learning

Instead of relying solely on spatial convolutions, the model integrates orthogonal transforms directly into the network layers to learn features in the frequency domain.

HT-UNet (Single Branch): A standard UNet backbone where convolutional blocks are interspersed with Hadamard Transform (HT) layers.
- Process: Input features undergo a 2D Hadamard Transform $\rightarrow$ Learnable scaling (element-wise multiplication with a weight matrix) $\rightarrow$ Soft-thresholding (suppressing small coefficients) $\rightarrow$ Inverse Hadamard Transform (IHT).
- Benefit: This creates an orthogonal latent space, forcing the network to focus on essential "frequency" components while reducing the need for heavy convolutional kernels.
TD-FusionUNet (Dual Branch): A scalable extension that fuses two parallel encoding branches:
- Upper Branch: Uses the Hadamard Transform (HT) for global mixing.
- Lower Branch: Uses the Discrete Cosine Transform (DCT) for frequency representation.
- Fusion: The outputs of both branches are upsampled, aligned via convolution, and fused via element-wise summation. This allows the model to leverage complementary strengths of both transform domains.

B. Custom Preprocessing Techniques

To address data sparsity and imbalance in fire masks, two specific techniques were introduced:

Random Margin Cropping: Uncertain pixels are perturbed. Background pixels (0) are replaced with small random values ( $U(0.01, 0.03)$ ), and fire pixels (1) with high random values ( $U(0.8, 0.99)$ ). This enriches the representation of binary masks and prevents overfitting to strict boundaries.
Gaussian Mixture Softening: A mixed Gaussian blur is applied to input masks using multiple standard deviations (e.g., $\sigma = \{0.4, 0.8\}$ ). The resulting blurred maps are averaged to create a "soft probability map," generating synthetic samples to complement missing or sparse data.

C. Training Strategy

Loss Function: A weighted combination of Binary Cross-Entropy (BCE), Dice Loss, and Focal Loss to handle class imbalance.
Datasets: Evaluated on the Google Next-Day Wildfire Spread dataset (2023) and the WildfireSpreadTS dataset.

3. Key Contributions

Lightweight Transform-Domain UNet: Introduced a novel architecture that replaces standard convolutional operations with trainable Hadamard and DCT layers, significantly reducing parameter counts while maintaining high representational capacity.
Dual-Branch Fusion: Demonstrated that fusing HT and DCT latent spaces improves feature extraction compared to single-transform or standard spatial-only models.
Robust Preprocessing Pipeline: Developed custom data augmentation strategies (Random Margin Cropping and Gaussian Mixture Softening) specifically tailored to handle the extreme sparsity of wildfire data.
Edge-Ready Design: The model is designed for deployment in resource-constrained environments (e.g., edge computing) due to its low parameter count and computational efficiency.

4. Experimental Results

The model was evaluated against baselines including 2D-CNN Autoencoders and ResNet18-based UNets.

Dataset	Model	Parameters	F1 Score	Precision	Recall	IoU
Google Dataset	2D-CNN Autoencoder (Custom)	78k	0.621	0.588	0.658	0.382
	HT-UNet (Custom)	45k	0.665	0.737	0.606	0.433
	TD-FusionUNet (Base=8)	93k	0.702	0.876	0.586	0.555
WildfireSpreadTS	Baseline UNet (ResNet18)	14M	0.589	0.536	0.654	-
	HT-UNet	169k	0.570	0.510	0.646	-
	TD-FusionUNet (Base=8)	370k	0.591	0.539	0.653	-

Efficiency: The TD-FusionUNet achieved an F1 score of 0.591 on the WildfireSpreadTS dataset using only 370k parameters, outperforming the ResNet18 baseline (14M parameters) which scored 0.589.
Performance: On the Google dataset, the TD-FusionUNet achieved the highest F1 score (0.702) and IoU (0.555), significantly outperforming both the CNN Autoencoder and the single-branch HT-UNet.
Scalability: The model showed consistent improvement as channel width increased, proving the gains were due to the transform-domain design rather than just parameter scaling.

5. Significance

Accuracy-Efficiency Trade-off: The paper demonstrates that orthogonal transform-domain learning can achieve state-of-the-art accuracy with a fraction of the computational cost of traditional deep learning models.
Real-Time Applicability: The lightweight nature (370k parameters) makes the model suitable for real-time wildfire prediction on edge devices, which is critical for rapid emergency response in remote or resource-limited areas.
Generalization: The proposed preprocessing techniques and transform-domain fusion provide a robust framework for handling sparse, imbalanced geospatial data, a common challenge in environmental monitoring.
Modularity: The transform layers are modular and can be integrated into other neural network architectures beyond UNet, offering a new direction for efficient feature learning in computer vision.