CrnE | Devpost

Inspiration

In an ever-changing world, farmers, businesses, and governments need accurate models to predict crop growth. Creating a low-cost, efficient tool to do so is, in our eyes, essential to keep up with modern technology, climate change, population growth, and more. This is where CrnE comes in.

What it does

CrnE takes in the last 3 years of weather data for a region, including average monthly temperature and precipitation totals, and outputs a prediction of the next year's corn yield, in tons/acre.

How we built it

We built our own custom Multi Layered Perceptron (MLP) model for this project using the PyTorch python library. We trained it on the last 50 years of temperature and precipitation data for over 30 states. Our model takes in 96 inputs, followed by one fully connected, linear hidden-layer with the same size as the input, followed by one fully connected, linear hidden-layer with half the size as the input, followed by one fully connected, linear hidden-layer with a fourth the size as the input, which connects to 1 final output. We used ReLU activation between each of the layers (except the output), as well as a dropout layer between our first and second hidden layers. The model was trained for ~5500 epochs with a batch size of 32, and we used an early stopping criteria of increasing loss for 3 epochs. We used MSELoss as our criterion and ADAM as our optimizer; both of these (especially MSELoss) are fairly common for regression networks.

Challenges we ran into

We had several teammates drop out of the project last-minute, including one who was going to do all of our front-end development; we ended up improvising and trying to have fun.