Using SSL to create useful representations from corrupted spectral-temporal data for applications in agriculture

Agriculture is central to questions of poverty, climate change, land use, and human and planetary health. Managing agricultural landscapes at regional and global scales necessitates equally large-scale data: satellite remote sensing imagery is a natural fit. Current methods in remote sensing have historically and are currently challenged by corrupted and missing data as well as the question of how to use the vast volumes of existing unlabeled data. We propose a new method and new data construct – an adaptation of Barlow Twins using the “d-pixel” – which we postulate will be effective for downstream tasks while both invariant to missing data and leveraging unlabeled data.