The Project:

You can find the FarmYeild Dataset on github, or you can clone it directly:

git clone https://github.com/raulium/FarmYield-Dataset.git

In the beginning

As a final project for my Computer Vision course, I wanted to perform two tasks:

  1. Predict crop yield, given satellite spectral imagery
  2. Classify crop type, given the same

In order to do this, I ended up having to build my own dataset in order to perform my work. I was put in contact with a farmer in Illinois that provided raw crop yeild data collected from his harvest over two years, which I parsed in CSV format.

The second half missing was the satellite imagery, which was a multistep process…

  1. Identify the spatial grid containing the farm, using lat/long and the Earth Explorer provided by the USGS.
  2. Download the appropriate images from Libra.
  3. Using a combination of GDAL and geoio, correlate the pixels of the imagery with the lat/long measurements provided to me in the yield CSV.

Then it was a matter of collating all the datapoints together into the overall dataset.

It’s been anonymized, but none of the lat/long data was important anyway for the purposes of classificaiton and crop yield.

Results

I wasn’t able to predict yield. It happens. I was, however, able to classify the crops as corn or soy beans with a degree of accuracy I was happy with.

You can read more in the paper.