{ "fetched_at": "2026-10-19T13:43:59Z", "source_url": "https://www.kaggle.com/competitions/2024-flame-ai-challenge/data", "pages": [ { "name": "data-description", "content": "## Files\n\n* **train.csv** - contains training set info\n* **test.csv** - contains test set info\n* **sample_submission.csv** - a sample submission file in the correct format generated via linear interpolation. See `generate_submission.ipynb` for how this is generated in the `Code` tab.\n* **data** - See `Starter_Notebook` in the `Code`tab on how to load the data and visualize it.\n\n\n## Columns\n* `id` - file id \n* `u` - wind speed\n* `alpha` - terrain slope\n* `Nt` - number of available snapshots (temporally correlated)\n* `Nx`, `Ny` - number of pixels in height and width.\n* `ustar_filename` - Filename for ustar velocity\n* `theta_filename` - Filename for potential temperature\n* `xi_filename` - Filename for fire location \n\n\n## Columns in sample_submission\n* `id` - file id \n* `0,1,...` - flattened array from fire location predictions\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F7329892%2Fd463685965e51e7bb515b704c62f2b7f%2Fflameai-fig-2.png?generation=1726962366515635&alt=media)", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "Participants in this Kaggle challenge will be provided with a high-fidelity dataset containing 2D wildfire flowfield data of wind speed (*ustar*), temperature (*theta*), and fire-line location (xi). To describe each case globally, we are also providing information about the slope (*alpha*) and mean wind speed (*u10*) -- both being scalar values, and are commonly used in the wildfire community to parameterize wildfire-spread behavior.\n\nYour task is to develop a state-of-the-art spatio-temporal ML model that can accurately forecast the 2D fire propagation as a function of time. The challenge consists in capturing the complex dynamics of wildfire spread that is affected by factors such as wind speed, terrain, and atmospheric turbulence. Your model should not only predict the instantaneous fire-line but also account for temporal variations, ensuring that forecasts remain robust and reliable over extended time periods. This will be critical for enhancing wildfire management and fire prevention strategies. To add interpretability, we encourage you to consider infusing physical information into your ML model. \n\n# Challenge\n\nWe are hosting seminars concurrently with the challenge. For more information, see our website https://flame-ai-workshop.github.io/\n\nThere are invited talks and seminars with experts in AI, ML, and physical flows. Join us to participate in them and ask questions. There is also a Slack channel with support to facilitate this challenge and encourage collaboration!\n\nRegister at https://events.humanitix.com/stanford-flame-ai-workshop-2024\n\n# Dataset\nThe dataset provided for this challenge contains high-fidelity 2D flowfield sequences of wildfire propagation that are extracted from 3D ensemble simulations. The sequences of flowfield data are represented as n channels (wind field, potential temperature, fire-line location, 2D grid, slope, and mean wind speed). The dataset will be divided into training and validation sets, and a separate test set will be used for evaluation. *This contains all the data available for this ML challenge, and you're **not** allow to use other data (we will retrain model submissions for verification).* \n\n## Key Features of the Dataset:\n\n1. **Spatial Resolution**: Each flowfield is represented as a 2D grid of data points, where each point contains information about key parameters such as velocity, potential temperature, wind speed, and fire location. The grid resolution captures both large-scale fire movements and finer details associated with turbulence, and fire-atmospheric coupling. Note that the fire-line location is binarized with 1 indicating the fire and 0 indicating no fire.\n\n2. **Temporal Data**: The dataset is temporal in nature, consisting of multiple sequential snapshots of fire propagation that are separated by equidistant intervals. Each sequence represents the evolution of a fire over time, allowing participants to model the time-dependent nature of the system.\n\n3. **Multiple Scenarios**: The dataset includes wildfire simulations under various conditions:\n- *Wind Conditions*: Simulations capture different mean wind speed, which significantly influence the fire-spread behavior. \n- *Terrain*: Data accounts for diverse terrains, from flat surfaces to high-slope conditions, affecting the fire-propagation speed.\n\n4. **Training and Testing Splits**: The dataset is pre-split into training and testing sets:\n- *Training Set*: Contains annotated flowfields with both input data (initial conditions, wind profiles, fire location).\n- *Testing Set*: Consists of input flowfields without corresponding fire spread labels. Participants will be required to predict fire propagation for these cases.\n\n5. **Data Format**: The dataset is provided in a standardized format, with each flowfield stored as a time-indexed sequence of 2D arrays in binary data format '.dat'. Accompanying metadata contain information about environmental conditions, including mean wind speed and terrain slope.\n\n# ML Task\nThe objective is to develop an auto-regressive ML model that predicts the fire-line location (xi) at a time t+1, using 2D information of wind field, potential temperature, and fire line at previous and current time ≤ t. While the primary evaluation focuses on your model’s predictions for the fire-line location, you are encouraged to leverage all the provided information for improved accuracy. For the leaderboard, your model will be evaluated based on its ability to forecast 20 timesteps on the testing dataset, starting from the end of the initial 5 steps provided.\n\n# Submission Requirements\n\n1. Use Kaggle's `Submit Predictions` to evaluate your generated predictions on the public leaderboard with MSE.\n2. Share a google drive: your final code [After Oct 17] using this google form: \nhttps://forms.gle/1v2SM9wPUoeXYxi87\n3. Share on google drive: trained weights and code for loading weights [After Oct. 17] using the same google form.\n\n# Getting Started\n\nTo get started, please review the provided dataset and the sample code provided in the Code subsection. You are encouraged to explore various deep learning architectures, data augmentation techniques, and pre-processing methods to achieve the best results. Please contact flame.ai.workshop@gmail.com for further information. ", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Evaluation", "content": "As a primary metric for model evaluation, we'll be using MSE(Ground Truth, Prediction) of the fire-line location to evaluate pixel-wise accuracy for the public leaderboard. The top-10 model submissions for invitation to join a journal submission will be reproduced and tested with additional evaluation metrics (SSIM, PSNR, error accumulation over time, robustness for long term prediction). [updated 10/9 -- blastnet] \n\n## Submission File\nFor each ID in the test set, you must predict a flattened (C-order) 113x32 fire line location. The file should contain a header and have the following format:\n\n id,0,1,2,...,72319\n 1,0,0,0,...,0\n 2,0,0,0,...,0\n 5,0,0,0,...,0\n 6,0,0,0,...,0\n etc.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "rules", "content": "1. Only use training data provided here.\n2. Open-source Pretrained Models are allowed.\n3. We will be checking and reproducing top code submissions at the end of the competition.\n4. Feel free to take the data offline and use your own platforms\n5. Be polite and Inclusive during discussions!\n6. code and solutions must be your own or your team’s work. Plagiarism, or presenting another’s work as your own, will lead to disqualification. ", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "abstract", "content": "**Welcome to the FLAME AI 2024 Challenge!**\nThe propagation of wildfires is a complex, dynamic process that is influenced by various factors, such as fuel, wind, terrain, and other environmental conditions. Accurately and reliably predicting the rate-of-spread of wildfires is of critical importance for fire management, rapid fire response, and fire mitigation.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F22578283%2F7e09b4a232ef2e70caec6b402336df3c%2Ffire_3d_u6s15Public.png?generation=1726969390383946&alt=media)\n\nIn this competition, you’ll have the opportunity to develop a spatio-temporal ML model to accurately forecast the fire propagation, considering effects of terrain, atmospheric flow conditions, and environmental interactions. The underlying data is extracted from a series of high-fidelity ensemble simulations that are the basis of the [Google FireBench](https://sites.research.google/gr/wildfires/firebench/) dataset, which consists of an ensemble of high-resolution large-eddy simulations that capture the three-dimensional wildfire-spread behavior and coupling with the atmospheric environment. \n\nBy considering a reduced data-set that was extracted from FireBench, your task is to create an autoregressive ML model that learns from these intricate flowfield conditions and generalizes to predict fire propagation in unseen conditions.\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Join the Community", "content": "We encourage participants to join the Kaggle and Slack (please register under https://flame-ai-workshop.github.io to be added to the slack channel) discussion forums for this challenge, where you can collaborate with other data scientists, ask questions, and share insights. \n\nReminder: There is a prize associated with collaborative activities!", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Timeline", "content": "1. Launching the Competition Page [Sept 23, after FLAMEAI launch event]\n2. Start of Challenge [9.30 am PDT, Sept 24]\n3. Python/ML Recap [9.30 am PDT, Sept 24]\n4. Kaggle Competitions Platform Tutorial [11.00 am PDT, Sept 24]\n5. Team Registration [11.59 pm, ~~Sept 30~~ Oct 11]\n6. Results Final Submission [11.59 pm PDT, Oct 17]\n7. Code Submission [11.59 pm PDT, ~~Oct 20~~ Oct 25]\n8. Prize Announcements [11.59 pm PDT, ~~Oct 24~~ Nov 1]", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Prizes", "content": "We are offering exciting prizes for:\n* 1st prize - $2500 per team\n* 2nd prize - $1500 per team\n* 3rd prize - $1000 per team\n\n\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Teams", "content": "1. Max of 2 people per team\n2. Can choose teammate from registered participants (Use Slack DM’s, #networking channel)\n3. Ask your friends if they’re interested (we can add them to slack channel)\n4. To form a team, we will provide a google forms link: https://forms.gle/pW3MSzdn2k1ptMWY6\n5. Deadline for team registration: Fri Sept 30, 11.59 pm PDT\n6. You may find the `Teams` tab convenient for team mergers for Kaggle team functionalities. ", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Citation", "content": "Matthias Ihme, Jen Zen Ho, Bassem Akoush, Wai Tong Chung. (2024). 2024 FLAME AI Challenge. Kaggle. https://kaggle.com/competitions/2024-flame-ai-challenge\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }