Programs
Open
Chase the Cloud Challenge '26 (ISRO)
Generative AI • Diffusion Models • Computer Vision • Satellite Analytics • Weather Intelligence
Evaluation only · Open
Build a Diffusion-Based AI system that predicts future cloud movement from past INSAT satellite images to enable accurate short-term weather forecasting.
Overview
Clouds can develop, move, and transform rapidly, but satellite images are available only at fixed intervals. Even a short gap in observation can make it difficult to track fast-changing weather systems. Predicting how clouds will move in the next few hours could support faster and more informed weather monitoring.
The brief
A weather team observes rapidly growing cloud formations that may develop into a thunderstorm. They have satellite images from the previous three hours, but the next image has not arrived yet. During this gap, the clouds may shift, intensify, or disappear. Traditional forecasting methods may struggle to generate realistic future satellite imagery from limited observations. Your challenge is to build a Diffusion-Based Cloud Motion Prediction System that learns from past INSAT satellite images and generates realistic future cloud frames for short-term weather nowcasting.
Dataset Requirement:
Participants should use meteorological imagery from the INSAT-3D, INSAT-3DR, or INSAT-3DS missions, accessed through the MOSDAC platform.
The INSAT imagers provide visible, shortwave infrared, mid-infrared, water-vapour, and thermal-infrared observations. Spatial resolution differs by channel: visible and SWIR imagery is nominally 1 km, MIR and TIR imagery 4 km, and water-vapour imagery 8 km at the sub-satellite point. (Mosdac)
Participants may select the geographical region, time period, channels, preprocessing method, and training sample size based on available computing resources. All data sources and preprocessing decisions must be properly documented.
Answer These Questions:
How can past satellite frames be used to understand cloud movement and development?
Can a diffusion model generate realistic future cloud images without losing important weather patterns?
How should predicted images be evaluated against actual satellite observations?
How can the solution support real-time thunderstorm, lightning, or severe-weather monitoring?
What To Build: Cloud Motion Prediction Model: Build a generative model that uses the previous three hours of satellite imagery to predict future cloud conditions.
Multi-Channel Satellite Processing: Process visible and infrared satellite channels to capture cloud structure, temperature, moisture, and movement.
Prediction & Comparison Dashboard: Display historical frames, generated predictions, actual observations, prediction errors, and cloud-motion visualizations.
Scalable Nowcasting Pipeline: Create a documented workflow for satellite-data preparation, model training, future-frame generation, and performance evaluation.
Prediction Goal
Use the previous six satellite frames covering three hours to generate: A minimum of two future frames, representing one hour ahead. Additional predictions extending up to six or seven hours, where technically feasible.
One Sentence
Complete this statement: “The moment my model successfully predicted where the clouds would move next was...”
Deliverables
- Prepared Multi-Channel Satellite Dataset
- Data Preprocessing and Frame-Sequencing Pipeline
- Diffusion-Based Cloud Prediction Model
- Predicted and Actual Frame Comparison
- Model Evaluation Report
- Interactive Dashboard or Visualization
- GitHub Repository
- 3–5 Minute Demonstration Video
Evaluation criteria
- Cloud Prediction Solution – 30% : Does the model generate meaningful and realistic future cloud frames from past satellite observations?
- Data Science & Technical Implementation – 30%: Are the satellite preprocessing, model architecture, training process, and evaluation methods technically sound?
- Forecast Quality & Practical Impact – 20%: Can the predictions support short-term weather monitoring, and are their limitations clearly explained?
- Documentation & Presentation – 20%: Are the code, methodology, visual comparisons, documentation, and demonstration clear and professionally presented?