OBSCURA: Satellite AI Research Challenge

about this event

OBSCURA | SATELLITE AI RESEARCH CHALLENGE — Observing Blocked Satellite Columns Using Remote-sensing and Auxiliary data. Can AI predict atmospheric pollution over industrial sites it has never seen?

Satellites measure the atmosphere every day, but clouds and retrieval limitations leave important gaps. OBSCURA challenges students to estimate five atmospheric measurements using Sentinel-2 satellite imagery, weather, terrain, and image-derived features.

OBSCURA is a student-led research challenge organized by BTAJI Crew, an AI and data science learning community. It combines rigorous machine learning with a real environmental monitoring problem, encouraging participants to experiment, validate their models, and communicate their research findings.

The challenge at a glance: predict five atmospheric targets simultaneously; inputs are Sentinel-2 satellite image chips and 29 numerical features; the dataset has 4,439 samples, approximately 9.19 GB; 3,218 training samples and 1,221 test samples; evaluation uses masked mean column-wise RMSE (MCRMSE), where lower is better; participation is open to eligible college and university students worldwide; individuals or teams of up to 20 students; initial prize pool of $50 USD plus digital participation certificates.

01 | THE RESEARCH CHALLENGE. Each sample represents an industrial facility during an observation window. Your task is to develop a machine learning model that predicts five atmospheric measurements using the provided imagery and environmental information. The targets are: t_ch4 (methane enhancement around the site), t_no2 (nitrogen dioxide enhancement), t_co (carbon monoxide enhancement), t_aai (smoke and dust, the absorbing aerosol index), and t_firms (thermal fire and flare activity).

Some training targets are blank because the corresponding satellite measurements were unavailable. A blank value is not zero. The evaluation ignores missing ground-truth measurements for each target while continuing to evaluate the available measurements.

What makes OBSCURA a research challenge? The training and test datasets contain different industrial facilities from different time periods. Models must generalize to new locations and environmental conditions rather than memorize familiar sites. Participants are encouraged to investigate different approaches, establish strong baselines, perform appropriate validation, compare experiments, and document their findings.

02 | HOW THE COMPETITION WORKS. OBSCURA uses Kaggle for the technical competition and Devpost for registration and final research documentation.

Step 1: Register on Devpost. Register for the OBSCURA hackathon on Devpost and review the eligibility requirements, rules, judging process, and prizes.

Step 2: Join the Kaggle Competition. Visit the official Kaggle competition, accept the participation rules, and explore the dataset and available resources. Read the Overview, Data, Evaluation, and Rules sections carefully before beginning.

Step 3: Develop Your Research Solution. Explore the starter notebooks, analyze the dataset, develop machine learning or deep learning models, perform experiments, and evaluate performance using appropriate validation techniques. You may use Kaggle notebooks and available computational resources, subject to Kaggle's usage limits and competition rules.

Step 4: Submit Predictions on Kaggle. Generate prediction files in the required CSV format and submit them through Kaggle. Kaggle manages the official technical evaluation and public and private leaderboards.

Step 5: Complete Your Devpost Research Submission. During the final submission week, complete your Devpost project and upload your research write-up as a PDF. Your research report should explain your methodology, data preparation, baseline models, experiments, validation strategy, results, findings, and conclusions. Include relevant Kaggle notebook or permitted GitHub repository links so your approach can be reviewed and reproduced. Research write-up guidance and a completed example are available in the Resources tab; participants may organize their reports differently, provided all required information is included. No participant presentation video is required.

Important: A valid Kaggle submission and a complete Devpost research submission are both required for prize and participation certificate eligibility.

03 | WHO CAN PARTICIPATE? OBSCURA welcomes eligible students from recognized colleges and universities worldwide, subject to the official age, geographic, and platform requirements. Participants may compete individually or in teams of up to 20 students. Students from different universities may collaborate within the same team, subject to the official team formation and competition rules. Team membership and participant information must be consistent across Kaggle and Devpost for eligibility verification.

04 | COMPETITION TIMELINE. All dates and times are in Pakistan Standard Time (PKT, GMT+5). Competition Launch: October 24, 2026, 8:00 PM. Kaggle Research Period: October 24 – November 24, 2026. Devpost Submissions Open: November 18, 2026. Final Submission Deadline: November 24, 2026, 8:00 PM. Winners Announced: November 28, 2026, 8:00 PM. Participants can work on their Kaggle models throughout the research period; the Devpost project submission window opens during the final week for research documentation and final project submissions.

05 | JUDGING AND EVALUATION. The official Kaggle private leaderboard determines the technical ranking using masked mean column-wise RMSE (MCRMSE). Lower scores indicate better prediction performance. Final prize eligibility is subject to verification by the organizing team, including: valid Kaggle submissions and final leaderboard results; a completed Devpost project and research PDF; clear documentation of methodology and experiments; reproducibility of the submitted solution where required; student eligibility, correct team information, and compliance with competition rules. The organizing team may disqualify entries that violate the competition rules or fail required verification checks.

06 | PRIZES AND CERTIFICATES. OBSCURA begins with an initial $50 USD cash prize pool: 1st Place $25 USD, 2nd Place $15 USD, 3rd Place $10 USD. Cash prizes are awarded per winning entry. An individual winner receives the corresponding prize; a winning team receives one prize for the team, not a separate cash award for each member. The prize pool may increase if additional sponsors or supporters join, and any changes will be announced officially.

Digital Certificates of Participation: every eligible student who completes the required submissions on both Kaggle and Devpost will receive an individual digital Certificate of Participation after verification. Winning a cash prize or achieving a specific leaderboard position is not necessary to receive a participation certificate.

07 | COMMUNITY AND SUPPORT. Join the BTAJI Crew WhatsApp Community to receive announcements, ask questions, discuss technical challenges, and connect with potential teammates. After joining, open the community groups list and select the Kaggle group for OBSCURA-related discussions and updates. Official dataset information, evaluation details, technical rules, and competition resources are available through the Kaggle competition page.

08 | FINAL NOTE. OBSCURA is more than a leaderboard competition. It is an opportunity to approach a real environmental monitoring problem through careful experimentation, scientific reasoning, and reproducible machine learning research. Beyond the clouds. Beyond familiar sites. Build AI that generalizes. Organized by BTAJI Crew.

tracks

when

tickets

prizes

faq