// machine learning
GitHub
released
WSTATTPixel-wise, crop classification deep learning model making use of transformers to analyze satellite data.PythonPyTorchNumPyJupyter
Project Details
As an extension to the state-of-the-art crop-classification model, STATT, I wanted to develop a model that made better use of sparse temporal data. To achieve this goal, I made the following additions to STATT's pipeline: added Transformers to each stage of the U-NET encoder to extract temporal context at each spatial resolution, used temporal pooling layers after each transformer that passed context vectors forward to the next layer to aggregate an attention score that better informs the result, and implemented multi-head attention to better model the many phenological cycles of the observed crops. In the end, when training with 6 timestamps across the year, my model increased its accuracy from 70% to 78%.
See Other Projects