Open robmarkcole opened 2 months ago
Interesting proof of concept. How well does this approach work for larger areas?
The paper that describes the work we did to create the dataset, 4-class models, and catchment-wide map time-series is here https://eartharxiv.org/repository/view/7527/
The segfomer models are published here https://zenodo.org/records/8172858
A full 12+ year, full catchment set of labeled data will be available soon. This consists of a trillions of labeled pixels and would serve as an amazing model training dataset
https://towardsdatascience.com/learn-transformer-fine-tuning-and-segment-anything-481c6c4ac802