// researchResearcher
former
AI4EarthAttended a 2 week research program where I created a pixel-wise crop classification model with transformers that takes in satellite data.Project Details
For 2 weeks, I was involved in an in-person research program through the University of Minnesota and National Science Foundation in Minneapolis. The experience consisted of several different machine learning projects from the lab of research Vipin Kumar. My group's chosen project centered on the state of the art crop-classification model, WSTATT. After brainstorming and debating the structure of our worktime, my group's first experiment was an ablation study on the inputs of weather data into the WSTATT model. It was found that variables pertaining to water (precipitation, vapor pressure, etc.) were more important to the model's predictions. The next experiment we carried out was the creation of a transformer-based alternative to STATT which retained its predictive power when less satellite timestamps were present. In the end, I served as the model's chief architect and the hypothesis was proven correct as the model retained nearly identical performance when observing only a quarter of the timestamps.
See Other Experiences