We noticed that corn crop yields in Illinois seem to be monotonically increasing with time, apart from some occasional, seemingly hard to envision dips. By looking at the resources and suggested papers, we realised these are very much related to Extreme Weather Disasters (EWD), particularly Droughts (summed up into various statistical indicators such as the Standard Precipitation Index (SPI)). Huge north-south divide led us to think we could cluster crop yields on a county-basis.
What it does
- PCA provides insights into differences among counties related to their geographical location.
- Temporal Fusion Transformer, an Attention-based Deep Neural Network, can effectively be used to predict future yield after seeing historical yields
How we built it
- Pytorch was used to build a TFT according to literature
Challenges we ran into
- Weirdly enough, Historical Damages (in USD) to crops do not improve the forecasting capabilities
What's next for import torch as tf
Probably some other hackathons! ;)