Crop Yield Prediction

Web App Homepage
Soil sensor metadata for a given plot of land
Scatterplot for seed application rate vs crop yield

Inspiration

Our group was inspired by John Deere's mission to maximize food production and minimize costs and externalities, and achieve this using the technologies at our disposal.

What it does

Our goal was to create a tool to help identify if certain pieces of land will produce large amounts of yield, to justify the resources and efforts placed behind it. We created a tool to help identify the estimated expected crop yield (in bushels/acre) based on various factors of a piece of land (moisture, seed application rate, elevation).

How we built it

We trained a gradient descent ML classification model based on soil moisture, seeding rates, and elevation data provided by John Deere, to predict crop yield for similar pieces of land in the Mid-West. We then created a React UI with a flask backend to visualize the crop yield estimates for a given piece of land and its soil and seeding properties.

Challenges we ran into

We ran into some initial challenges with training the model, as we identified that the "Wet Mass" property directed correlated with the Yield value, which was overfitted and did not provide any new insight into the yield. We fixed this by finding new data that be used to better predict Yield values. Another technical challenge was in developing front-end since neither of our group members were comfortable working on it. We fixed this by meeting with the John Deere team during the mentor session for help in understand the front end for the React UI.

Accomplishments that we're proud of

Parsing datasets and tuning our model to achieve an accuracy of 87.42% Make a user friendly UI using CSS and ReactJS

What we learned

We were able to meet with many John Deere mentors, and understand the problems that the agricultural technology industry is facing today, such as water contamination from pesticides and the high costs involved with purchasing pesticide and herbicide. They were able to guide us in the right directions, which helped us gain a stronger understanding of our mission and product. We were amazed by the tremendous work that goes behind agriculture, and are happy to be able to create a product to help with the mission!

What's next for Crop Yield Prediction

Being able to empower farmers further by using historical weather data and soil conditions to predict for future expected crop yield along with other factors Use better soil data provided by the US census Factor in nearby streams into the model