🤖 machine learning

Spatio-Temporal Wildfire Spread Prediction in Canada using a Video Swin-Hybrid-U-Net and Satellite Imagery

This study proposes a Video Swin-Hybrid-U-Net deep learning framework that leverages publicly available satellite imagery and meteorological data to accurately predict spatio-temporal wildfire spread patterns across Canada.

Original authors: Maulik Srivastava, Esha Saha, Hao Wang

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Maulik Srivastava, Esha Saha, Hao Wang

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 trying to predict how a wildfire will move across the Canadian landscape tomorrow. It's like trying to guess the path of a chaotic, living river of fire that is constantly changing shape, speed, and direction based on the wind, the dryness of the ground, and the steepness of the hills.

This paper introduces a new "crystal ball" for firefighters and scientists: a smart computer program called Video Swin-Hybrid-U-Net. Here is how it works, explained simply:

The Problem: Old Maps vs. Moving Movies

Traditional ways of predicting fires are like looking at a single, static photograph of a landscape and guessing where the fire might go. They often miss the "movie" aspect—how the wind shifts hour-by-hour or how the ground dries out during the day. Other computer models are like watching a slideshow of photos; they see the pictures, but they don't really understand the motion between them.

The Solution: Teaching the Computer to Watch a Movie

The authors built a system that treats wildfire prediction like watching a video. Instead of just looking at one day's weather, the model looks at a three-day "movie" of the environment.

The Input (The Script): The computer watches a 3-day sequence of "frames." Each frame is a map containing:
- Where the fire was (from satellite photos).
- The weather (temperature, rain, wind speed, and wind direction), sampled every 6 hours to catch daily changes.
- The ground (how wet the soil is).
- The landscape (how steep the hills are and what kind of plants are there).
The Engine (The Director): The core of the system is a special type of AI called a Video Swin Transformer. Think of this as a director who doesn't just look at one actor (a single pixel on a map) but watches how the entire scene changes over time. It pays attention to the "long-range" connections—how a gust of wind in the morning might push the fire to a specific hill three days later.
The Output (The Forecast): After watching the three-day movie, the AI draws a single, clear picture of where the fire is likely to be tomorrow.

How It Learned (The Training Camp)

The model was trained on a massive library of real Canadian wildfires from 2014 to 2023. The researchers used only free, public data from Google Earth Engine, making the whole process transparent and open to everyone.

They taught the model by showing it thousands of examples of fires and asking, "Where will this fire be tomorrow?" The model made mistakes, corrected them, and learned the patterns of fire behavior specific to Canada's vast forests.

The Results: A Sharper Eye

When they tested this new "movie-watching" AI against older models (which were like looking at still photos or simple slideshows), the results were impressive:

Better Accuracy: The new model was significantly better at guessing the exact shape and location of the fire. It caught about 65% of the correct fire areas, while the next-best model only caught about 60%, and older models struggled around 36-45%.
Handling Steep Hills: The model got even better at predicting fires on steep slopes. It seems that fires moving up steep hills follow more predictable paths, and the AI learned to spot these patterns easily.
Weather Independence: Interestingly, the model remained accurate whether the ground was very dry or moderately wet. It learned to adjust its predictions based on the fuel conditions.

Why This Matters

The paper claims this is a major step forward because it treats time as a real dimension, not just a list of numbers. By understanding the sequence of events (like a video) rather than just a snapshot, the model captures the complex, dynamic nature of wildfires in Canada.

The authors conclude that this framework provides a powerful, open-source tool that can help authorities allocate resources and protect communities, but they emphasize that the model's true strength lies in its ability to learn from the "video" of the environment rather than just static maps.

Technical Summary: Spatio-Temporal Wildfire Spread Prediction in Canada using a Video Swin-Hybrid-U-Net

Problem Statement
Wildfire regimes in Canada are escalating in frequency and intensity, posing severe threats to ecosystems and infrastructure. While traditional physics-based simulators (e.g., FARSITE) offer interpretability, they often require extensive parameter tuning and are computationally intensive. Conversely, existing deep learning approaches frequently lack scalability or fail to effectively capture the temporal dynamics of fire spread, particularly within the unique environmental context of Canada's boreal forests. Furthermore, many advanced models rely on proprietary data, hindering reproducibility and the development of transparent, operational tools. The core challenge addressed is the need for a scalable, deep learning framework capable of modeling complex spatio-temporal dependencies in environmental data to forecast next-day fire incidence accurately.

Methodology
The study proposes a novel deep learning framework that frames wildfire spread prediction as a video-to-image semantic segmentation task. The model ingests a three-day sequence of multi-channel environmental data to predict a binary fire mask for the subsequent day.

Data Pipeline: The dataset is curated exclusively from public repositories via Google Earth Engine (GEE), covering major Canadian wildfire events from 2014 to 2023. Data is processed to a 500m spatial resolution (224x224 pixels).
- Dynamic Variables (Time-varying): Includes Fire Masks (MODIS), Weather Data (ECMWF/ERA5-LAND: temperature, precipitation, wind components sampled at 6-hour intervals), and Soil Moisture (GLDAS). These provide high temporal resolution to capture diurnal patterns.
- Static Variables (Time-invariant): Includes Topography (Elevation, Slope, Aspect from NRCan/CDEM), Vegetation (NDVI from NOAA/VIIRS), and Population Density (CIESIN/GPW). These are replicated across the three-day input sequence.
- Input Tensor: The data is structured as a 3D tensor $(T, C, H, W)$ where $T=3$ days, $C=23$ channels (18 dynamic + 5 static), and $H, W=224$ .
Architecture (Video Swin-Hybrid-U-Net):
- Spatio-Temporal Encoder: Utilizes a 3D Swin Transformer backbone. Unlike models that flatten time into channels, this encoder processes the input as a unified 3D volume. It employs shifted 3D window attention (SW-MSA) to efficiently model relationships across both space and time simultaneously, capturing long-range dependencies and dynamic evolution.
- Convolutional Decoder: A U-Net style decoder uses 3D transposed convolutions to upsample features to the original resolution. It integrates skip connections from the encoder to recover fine-grained spatial details. The final 3D feature tensor is aggregated along the temporal dimension via averaging to produce a 2D segmentation mask.
- Loss Function: To address severe class imbalance (fire pixels vs. non-fire pixels), the model is trained using a weighted sum of Binary Cross-Entropy (BCE) and Dice Loss ( $\lambda_1 = \lambda_2 = 0.5$ ).
Training: The model was trained end-to-end using the AdamW optimizer with a Cosine Annealing Warm Restarts schedule. Experiments were repeated five times with different random seeds to ensure robustness.

Key Contributions

Novel Architecture with Spatio-Temporal Attention: The paper introduces a Video Swin-U-Net architecture that treats input data as a 3D video volume. This design departs from 2D convolutional or spatial-only attention models, enabling the explicit learning of complex, long-range dependencies across space and time.
Deep Temporal Feature Learning: The model leverages a three-day sequence with high temporal granularity (variables sampled four times daily). This allows the capture of transient diurnal patterns in meteorological variables, which are critical drivers of fire behavior but often missed by models using daily inputs.
Publicly Accessible Framework: The entire pipeline relies on open-source data from GEE, ensuring the methodology is transparent, reproducible, and scalable for a Canada-wide operational tool, addressing the barrier of proprietary data in existing systems.

Results
The proposed Hybrid Video Swin-U-Net was evaluated against three baselines: a standard 2D U-Net, a 2D Swin U-Net, and a pure Video Swin U-Net (using 3D blocks in both encoder and decoder).

Quantitative Performance: The Hybrid model achieved superior performance across all metrics on the held-out test set:
- F1-Score: 0.6550 (vs. 0.6003 for the next-best Video Swin U-Net and 0.4528 for the 2D Swin U-Net).
- Precision: 0.6328.
- Recall: 0.6791.
Topographical Analysis: Performance varied with terrain. The model showed a strong positive correlation between slope steepness and accuracy (F1-Score increasing from ~0.62 on flat terrain to ~0.72 on steep terrain). Elevation showed a non-linear relationship, with higher accuracy at low and high extremes but a dip in mid-range elevations.
Soil Moisture Analysis: The model demonstrated robustness across varying soil moisture conditions, maintaining consistent F1-Scores (0.64–0.68) across different moisture bins for all three input days, indicating effective learning of fuel availability dynamics.
Qualitative Analysis: Visual inspection revealed the model's ability to capture complex fire shapes and boundaries, producing smoother contours compared to the jagged ground truth masks.

Significance and Claims
The paper claims that the explicit use of spatio-temporal attention mechanisms is paramount for dynamic environmental forecasting. By processing time as a distinct dimension rather than a flattened channel, the model overcomes the limitations of 2D approaches in distinguishing temporal sequencing from spatial features. The study demonstrates that modern computer vision architectures, specifically those adapted for video data, can significantly advance the prediction of environmental hazards like wildfires.

The authors position this work as a step forward in applying state-of-the-art techniques to Canadian wildfire dynamics, offering a transparent and reproducible framework that captures the unique interplay of weather, topography, and fuel. The results suggest that such models can provide accurate, spatially explicit forecasts essential for resource allocation and evacuation planning. The paper concludes modestly, noting limitations such as dependence on optical satellite data (susceptible to cloud cover) and the redundancy of replicating static channels, while suggesting future directions like longer prediction horizons and improved static data integration.