OBSCURA | PARTICIPANT RESOURCES

Everything you need to understand the research challenge, explore the dataset, develop your models, submit predictions, and prepare your final research write-up.

01 | WATCH THE VIDEO GUIDES

We recommend watching both videos to understand how OBSCURA works and how to get started.

Video 1: OBSCURA Introductory Walkthrough

This video explains:

• How to register and navigate the Devpost hackathon.
• How to join the Kaggle competition.
• Where to find datasets, notebooks, leaderboards, and discussions.
• Individual participation and team formation.
• How Kaggle and Devpost submissions work together.
• Important resources, deadlines, and submission requirements.

Watch the Introductory Walkthrough

Video 2: Understanding the OBSCURA Research Problem

This video provides a more detailed explanation of:

• The research problem and scientific motivation.
• Satellite imagery and environmental features.
• The five atmospheric prediction targets.
• Dataset organization and missing target values.
• Baseline approaches and model development.
• Evaluation, validation, and generalization.

Watch the Detailed Research Explanation

02 | DOWNLOAD AND EXPLORE THE DATA

Access the official dataset from the Kaggle Data section.

The dataset contains approximately 4,439 samples (9.19 GB), including:

• train.csv: 3,218 training samples with numerical features, anonymized site information, and available target values.
• test.csv: 1,221 test samples containing features for prediction.
• sample_submission.csv: Required prediction submission format.
• images/: 4,439 multispectral Sentinel-2 image chips.
• README.md: Additional dataset information and instructions.

Each satellite image chip contains 14 bands, and the provided tabular data includes 29 numerical features.

Important: Do not redistribute restricted competition data, image chips, target files, or identifying information outside authorized channels.

03 | OFFICIAL STARTER NOTEBOOK

Our Kaggle starter baseline notebook is available to help participants understand the technical workflow.

Open the OBSCURA Starter Baseline Notebook

Use the notebook as a starting point for:

• Loading and exploring competition data.
• Understanding numerical inputs and prediction targets.
• Handling missing target values.
• Implementing site-grouped validation.
• Training baseline machine learning models.
• Evaluating model performance.
• Generating Kaggle prediction submissions.

Participants are encouraged to build upon the baseline, investigate alternative methods, and document their experiments.

Kaggle notebooks may provide CPU and GPU resources, subject to platform availability, account requirements, and usage quotas.

04 | RECOMMENDED RESEARCH WORKFLOW

1. Review the official Kaggle Overview, Data, Evaluation, and Rules sections.
2. Watch the introductory and detailed research explanation videos.
3. Explore the dataset and official baseline notebook.
4. Study the 29 numerical features and 14-band satellite image chips.
5. Understand the five prediction targets: t_ch4, t_no2, t_co, t_aai, t_firms.
6. Validate using site-grouped splits based on site_key.
7. Handle missing targets appropriately during training and evaluation.
8. Develop baseline models and test improvements systematically.
9. Evaluate experiments using the official masked mean column-wise RMSE metric.
10. Submit predictions through Kaggle and document your research for Devpost.

Individual students and teams of up to 20 eligible students may participate, subject to the official rules.

05 | RESEARCH WRITE-UP RESOURCES

Each individual entry or team must upload a research write-up PDF as part of the final Devpost project submission.

We have provided two documents to help you prepare.

Document 1: Research Write-up Guidance

This document explains what information to include in your research report, such as methods, baseline experiments, validation, results, limitations, and lessons learned.

View Research Write-up Guidance PDF

Document 2: Completed Research Write-up Example

This document demonstrates how a completed research report can be organized and presented.

View Completed Research Example PDF

Both documents are also available in our Research Resources Folder.

These resources are examples and guidance, not mandatory report templates.

You may use your own report structure, include additional experiments, combine sections, and present your methodology in a way that best communicates your research.

Your final PDF should clearly explain your problem understanding, methodology, experiments, validation, results, findings, and relevant code or notebook references.

A participant presentation video is not required.

06 | FINAL SUBMISSION CHECKLIST

Devpost Submission Window: November 18 - November 24, 2026

Final Deadline: November 24, 2026, at 8:00 PM PKT

• Complete at least one valid Kaggle prediction submission.
• Ensure your Kaggle profile and team information are correct.
• Create your corresponding Devpost project.
• Upload the required research write-up PDF.
• Include relevant permitted Kaggle notebook or GitHub repository references.
• Ensure the participating team members match across both platforms.
• Follow the competition's eligibility, data-use, and submission rules.
• Complete all required submissions before the deadline.

Both Kaggle and Devpost submissions are required for cash prize and participation certificate eligibility.

07 | COMMUNITY AND SUPPORT

For announcements, questions, technical discussions, and team formation, join the BTAJI Crew WhatsApp Community.

Join Our WhatsApp Community

Once you've joined, open the community's groups list and select the Kaggle group for OBSCURA-related updates and discussions.

For technical competition questions, you may also use the official Kaggle discussion forum. Report sensitive data integrity or leakage concerns privately to the organizers.

Beyond the clouds. Beyond familiar sites. Build AI that generalizes.