Kernl

Farmers make high stakes decisions before they know how the season will unfold. Seed selection, planting rates, and input spending all depend on weather that has not happened yet. We were motivated by the idea that near term forecasts could be translated into practical guidance. Instead of only predicting yield, we wanted to help growers understand whether they chose the right corn or soy variety given the season ahead.

Kernl evaluates corn and soy varieties under the current 60 and 90 day forecast. It projects expected yield for each variety and displays an uncertainty band so farmers can see their risk exposure. Stress sensitive varieties show wider ranges under heat or drought conditions, while more tolerant varieties show tighter, more stable outcomes. We also translate yield scenarios into revenue ranges so growers can see how weather risk connects directly to their bottom line.

To build this, we combined historical county level yield data with daily weather inputs and forecast data. We engineered features such as heat stress days, rainfall deficits, and growing degree days, then trained regression models in Databricks to predict yield and estimate uncertainty. We tracked experiments, compared model performance, and built comparison views that allow corn and soy varieties to be evaluated under the same forecast conditions.

Along the way, we dealt with permission limits in Databricks and had to adjust how we stored and managed data. We also worked within runtime constraints and avoided installing unsupported packages. Aligning historical weather and yield data required careful cleaning and feature design, but it strengthened our pipeline.

We are proud that we built a working yield prediction system with strong performance and interpretable outputs. The uncertainty bands make risk visible instead of hiding it behind a single average prediction. The variety comparison view clearly shows how seed choice changes expected yield and volatility under the same forecast.

Through this process, we learned that uncertainty is just as important as prediction. Farmers care about downside risk and revenue impact, not just averages. We also gained experience building an end to end workflow in Databricks, from ingestion and feature engineering to modeling and evaluation.

Next, we plan to incorporate hybrid level trial data and more granular field level weather inputs. We also want to build a live what if engine that allows users to adjust temperature or rainfall assumptions and immediately see changes in projected yield and revenue. Our goal is to develop Kernl into a full decision support tool that supports farmers from seed purchase through harvest planning.

Built With

• databricks