ICLR Workshop Challenge #2: Radiant Earth Computer Vision for Crop Detection from Satellite Imagery
$5,000 USD
Identify crop type using satellite imagery, and win a trip to present your work at ICLR 2020 in Addis Ababa.
468 data scientists enrolled, 110 on the leaderboard
AgricultureComputer VisionSatelliteUnstructuredSDG2
3 February 2020—29 March 2020
55 days

The objective of this competition is to create a machine learning model to classify fields by crop type from images collected during the growing season by the Sentinel-2 satellite.

The dataset contains a total of more than 4,000 field images, which have been split into training and test sets (3,286 in the train and 1,402 in the test). You will train your machine learning model on the fields included in the training set and will apply your model to the fields in the test set. You will submit your predictions for the crop type for each of the fields in the test dataset.

The data you will have access to includes 12 bands of observations from Sentinel-2 L2A product (observations in the ultra-blue, blue, green, red; visible and near-infrared (VNIR); and short wave infrared (SWIR) spectra), as well as a cloud probability layer. The bands are mapped to a common 10mx10m spatial resolution grid.

The dataset can be downloaded in four tiles. These tiles are smaller than the original Sentinel-2 tile that has been clipped and chipped to the geographical area that labels have been collected.

Each tile has:

  • 13 multi-band observations throughout the growing season. Each observation includes 12 bands from Sentinel-2 L2A product, and a cloud probability layer. The twelve bands are [B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B11, B12] (refer to Sentinel-2 documentation for more information about the bands). The cloud probability layer is a product of the Sentinel-2 atmospheric correction algorithm (Sen2Cor) and provides an estimated cloud probability (0-100%) per pixel. All of the bands are mapped to a common 10 m spatial resolution grid.
  • A raster layer indicating the crop ID for the fields in the training set.
  • A raster layer indicating field IDs for the fields (both training and test sets). Fields with a crop ID 0 are the test fields.

Read more about raster images here.

The dataset for this competition is publicly available on Radiant MLHub. In order to access the data, you need to sign up for an API access token here. Then you can access the data using Radiant MLHub API (docs available here). The multi-spectral dataset is named “ref_african_crops_kenya_02_source” and the labels and field IDs dataset is named “ref_african_crops_kenya_02_labels”. You can use these example notebooks to access, download and load the data into your Python environment. Models that use the Field ID as a feature will not be accepted as a winning solution.

Your submission needs to be in the form of SampleSub.csv.

Here is a summary of the different classes.

Crop ID   Crop Type                  Number of fields in training data 
   1      Maize                                      1462   
   2      Cassava                                     829     
   3      Common Bean                                  98   
   4      Maize & Common Bean (intercropping)         487   
   5      Maize & Cassava (intercropping)             172   
   6      Maize & Soybean (intercropping)             160  
   7      Cassava & Common Bean (intercropping)        78