🚀 Bhartiya Antriksh Hackathon (BAH) 2026

Problem Statement: Infrared Image Colorization and Enhancement for Improved Object Interpretation

Welcome to the Bhartiya Antriksh Hackathon 2026! This repository provides a baseline implementation and technical guidelines for the challenge of transforming raw Thermal Infrared (TIR) satellite imagery into interpretable, colorized visual representations.

---

The Challenge

Thermal Infrared (TIR) data is invaluable for monitoring wildfires, urban heat islands, and volcanic activity. However, raw TIR imagery is typically single-band (grayscale) and lacks the intuitive detail of RGB imagery, making object interpretation difficult for human analysts.

Your Goal: Develop a computational pipeline and machine learning model that produces two primary outputs:

1. A Super-Resolved TIR Image: Increase the spatial resolution of raw TIR imagery to recover critical structural details.
2. A Colorized TIR Image: Synthesize realistic colors for the TIR data, using multi-spectral RGB data as a guide.

---

Data Acquisition

Data Specifications

On the USGS Earth Explorer site, all Landsat 9 bands (B2, B3, B4, and B10) are registered and provided at a 30m resolution. However, it is important to note that the original spatial resolution of the TIR band (B10) is 100m, while the RGB bands (B2, B3, B4) are natively 30m.

To get started, you will need Landsat 9 imagery.

Quick Start (Demo Data)

Use the provided bash script to download sample bands into input/demo_product/:

chmod +x scripts/download_data.sh
./scripts/download_data.sh

Custom Downloads (Google Earth Engine)

Use scripts/download.py to fetch specific bands using GEE:

python scripts/download.py <product_id> <bands> <start_date> <end_date> <output_path> --ee_project_id <your_project_id>

Custom Downloads (USGS Earth Explorer)

You may also download data directly from USGS Earth Explorer; please ensure it is placed in the input directory following the structure below.

Required Input Directory Structure

To ensure the baseline scripts function correctly, please organize your data as follows:

input/
└── <folder_name>/
    ├── <file_prefix>_B10.TIF
    ├── <file_prefix>_B2.TIF
    ├── <file_prefix>_B3.TIF
    └── <file_prefix>_B4.TIF

Note: While <folder_name> can be any identifier of your choice, the files inside must end with the specified band suffixes (_B10.TIF, _B2.TIF, _B3.TIF, _B4.TIF) to be correctly processed by the pipeline.

---

Baseline Implementation: Dataset Generation

This baseline focuses on the most critical part of the pipeline: creating co-registered training pairs.

Dataset Generation Workflow

Run the driver script to generate multi-resolution, spatially aligned patches:

python driver.py

Pipeline Details:

1. Merge: Optical bands (B2, B3, B4) are merged into a 30m RGB image.
2. Downscale: The baseline takes the 30m resampled USGS data and downscales it to create training pairs:
   • Input: All bands are processed from their 30m resampled versions.
   • Rescaling Factors:
     • RGB (30m) $\xrightarrow{\times 3.33}$ 100m
     • TIR (30m) $\xrightarrow{\times 3.33}$ 100m
     • TIR (30m) $\xrightarrow{\times 6.67}$ 200m
   • Data Flow: For the Super-Resolution task, the TIR band is downsampled to 200m as input, with the objective of recovering a 100m output.
3. Extract Co-registered Patches:
   • SR Pair: 256x256 (200m TIR) $\rightarrow$ 512x512 (100m TIR).
   • Colorization Pair: 256x256 (100m TIR) $\rightarrow$ 256x256 (100m RGB).
4. Save Output: Both .npy (for training) and .png (for verification) files are saved in output/patches/.
   • ⚠️ Important: Do not train your models on the .png files. .png files are intended for visualization purposes only. For training, use the .npy files to maintain the original radiometric resolution of the data.

Technical Alignment

The baseline ensures strict spatial co-registration:

• One pixel in the 200m TIR image corresponds exactly to a 2x2 block in the 100m TIR/RGB images.
• All patches are extracted using the same top-left offset to maintain alignment across resolutions.

---

Conceptual Workflow

The following diagram illustrates the end-to-end process for dataset generation. While we provide a baseline, these are suggested approaches. You are encouraged to explore alternative workflows for better structural accuracy or design entirely new pipelines to achieve the objectives.

graph TD
    A[Start: Raw Data Source] --> B(Download Landsat 9 Bands: B2, B3, B4, B10 - 30m)
    
    B --> C1(Merge B2, B3, B4 into RGB Image - 30m)
    B --> C2(Downscale TIR B10 by 3.33x - 100m)
    B --> C3(Downscale TIR B10 by 6.67x - 200m)
    
    C1 --> D(Downscale RGB by 3.33x - 100m)
    
    D --> E1(Create Image Patches: 100m RGB & 100m TIR)
    C2 --> E2(Create Image Patches: 100m TIR & 200m TIR)
    C3 --> E2

Loading

Expected Pipeline and Output Format

Following the dataset generation, you are expected to implement a multi-stage model pipeline.

Inference Flow: For the entire pipeline during inference, the input will be the 200m resolution TIR band (B10). The pipeline is expected to produce the two outputs detailed below.

1. Super-Resolution Stage: Develop a model to generate high-resolution (100m) TIR images from the low-resolution (200m) inputs.
2. Colorization Stage: Pass the resulting high-resolution TIR images into a colorization model to produce synthetic, interpretable RGB representations.

Mandatory Output Structure

To ensure standardized evaluation, your final output must be organized in the output/ directory as follows:

output/
└── model_outputs/
    ├── tir_superresolved_100m/
    │   └── <product_id>.tif
    └── colorized_tir_100m/
        └── <product_id>.tif

Note: <product_id> must exactly match the original input product ID.

Band Ordering Requirement: For the colorized TIR images, the output TIFF must adhere to the following channel sequence:

• Layer 1: Blue
• Layer 2: Green
• Layer 3: Red

---

Submission Guidelines

Required Deliverables:

1. Codebase: A link to your GitHub repository.
2. Model Weights: Your trained model weights (e.g., .pth, .h5).
3. Technical Report: A PDF detailing your approach and results.
4. Sample Results: A sequence of Raw TIR $\rightarrow$ Super-Resolved TIR $\rightarrow$ Colorized TIR.