Research · AI project2023.09 – 2023.12Team of 4
IAB capstone — Wildfire spread prediction with spatiotemporal deep learning
A U-LSTM that predicts the next wildfire front from the current front, weather and terrain
Problem
Wildfires spread quickly with wind and vegetation, yet Korea had no real-time spread prediction model
My role
Design and train the spatiotemporal model that predicts where the fire will be next
What I did
- Rasterized fire fronts from Korea Forest Service maps plus wind, NDVI and elevation in ArcGIS for four large fires (3 for training, 1 held out)
- Designed a shallow U-LSTM (2,736 parameters) to fit the small dataset (4 fires, 35 time steps)
- Expanded the sequences with sliding windows (length 1–19) into 151 training batches; evaluated with Overall Accuracy and F1
Results
- Unseen fire (Gangneung): Accuracy 0.90 · F1 0.85 (validation 0.95 · 0.81)
- Refactored the team notebooks into a tested Python package and documented the limitations on GitHub
Decisions & trade-offs
What I chose, and what I gave up for it
Only four fires of data: instead of a bigger model, a shallow network with 2,736 parameters, and sliding windows of length 1–19 to grow the training set to 151 batches.
For a fair generalization check, one entire fire (Gangneung) was held out for testing — less training data, but a real 'unseen fire' result.
Looking back · open questions
What I would change, and what is still open
I didn't trace which axis the tensors actually flowed along; I later realized the LSTM was looking at the spatial axis rather than time.
If I rebuilt it: ConvLSTM, event-level cross-validation, Dice loss + IoU, and satellite-based fire-front data.