# Competition: Beyond Visible Spectrum - AI for Agriculture 2026

## Overview
Participants are challenged to develop robust **Self-Supervised Learning (SSL)** methods to extract meaningful representations from satellite imagery, which will then be fine-tuned for **Crop Disease Prediction**.

Innovative approaches that leverage **Multimodal** data are highly encouraged to improve classification accuracy under varying environmental conditions.

## 1. Self-Supervised Training Data (Unlabeled)
The core dataset consists of unlabeled Sentinel-2 (S2A) time-series imagery, organized by location and acquisition time. This data is intended for pre-training SSL models (e.g., MoCo, SimCLR, MAE) to learn temporal and spectral features of agricultural land.

**Download:** https://www.dropbox.com/scl/fi/d7cmjdd10qlkp9w895me2/s2a.tar.gz?rlkey=xk1pka3q4nokpmj7dilfys5lw&st=xn8g7k53&dl=0)

If you encounter issues downloading the full `s2a.tar.gz` file due to network instability or size limits, you can download the dataset in split parts using the links below.

We have updated the link.

Please download all the following parts:

- [s2a.tar.gz.partaa](https://www.dropbox.com/scl/fo/0ca5g354y1tn8cw7dwk85/AITOGzyxyqCoKJjhzXbuOH4/s2a.tar.gz.partaa?rlkey=h8yirspxk6r312634qabi412m&st=x90erzjv&dl=0)
- [s2a.tar.gz.partab](https://www.dropbox.com/scl/fo/0ca5g354y1tn8cw7dwk85/AN5iLpgIOwJ6TWNhqAzuDcw/s2a.tar.gz.partab?rlkey=h8yirspxk6r312634qabi412m&st=v6m8n9w7&dl=0)
- [s2a.tar.gz.partac](https://www.dropbox.com/scl/fo/0ca5g354y1tn8cw7dwk85/AJutAYqAJ8hQ69aIL4RXZyQ/s2a.tar.gz.partac?rlkey=h8yirspxk6r312634qabi412m&st=ed7eqamq&dl=0)
- [s2a.tar.gz.partad](https://www.dropbox.com/scl/fo/0ca5g354y1tn8cw7dwk85/AP0sa-5mR8mzBoYBN10sa9w/s2a.tar.gz.partad?rlkey=h8yirspxk6r312634qabi412m&st=p0bii2m4&dl=0)
- [s2a.tar.gz.partae](https://www.dropbox.com/scl/fo/0ca5g354y1tn8cw7dwk85/AG3d3HP7pOgePdY4ni2yJ9s/s2a.tar.gz.partae?rlkey=h8yirspxk6r312634qabi412m&st=odnubqp0&dl=0)


### Data Structure
The data is organized hierarchically:
Band_File.tif

*   **Band Files**: Individual GeoTIFF files for each spectral band.

### Available Bands (Sentinel-2)
Each acquisition folder contains the following 12 spectral bands:
*   **Coastal/Aerosol**: `B1.tif`
*   **Visible**: `B2.tif` (Blue), `B3.tif` (Green), `B4.tif` (Red)
*   **Red Edge**: `B5.tif`, `B6.tif`, `B7.tif`
*   **NIR (Near Infrared)**: `B8.tif`, `B8A.tif`
*   **Water Vapour**: `B9.tif`
*   **SWIR (Short-Wave Infrared)**: `B11.tif`, `B12.tif`


## 2. Supervised Training Data (Labeled)
This dataset contains labeled samples for the downstream classification task.

##Submission Columns

*   `Id` - file name
*   `Category` - Aphid, Rust, RPH, Blast