Barlow Twins representations and AGB maps examples for Pará state, Brazil

Clement Atzberger, Markus Immitzer, Josue Lopez Ruiz

Published: 31 Mar 2025, Last Modified: 04 Nov 2025ZenodoEveryoneRevisionsCC BY-SA 4.0
Abstract: Barlow Twins representations for S2 tile 22MDU Barlow Twins were originally developed for the processing of natural (RGB) images (Zbontar et al., 2021). The spectral-temporal Barlow Twins (Lisaius et al., 2024), on the other hand, were developed to derive representations from pixel-wise spectral-temporal observations. The approach is fully self-supervised and does not use any inputs other than the multi-spectral time series, together with the corresponding cloud masks. For more details, the reader is referred to Atzberger et al. (2025, A scalable, annual aboveground biomass product for monitoring carbon impacts of ecosystem restoration projects, Remote Sensing of Environment). Years 2013-2016 are based on Landsat, 2017-2024 on Sentinel-2 Naming: representation_YEARModelApplied_S2tile_bt_YearBarlowTwinModelTrained_u8.tif, e.g., representations_2024_22MDU_bt_2019_u8.tif AGB maps for S2 tile 22MDU A simple, 6-layer fully connected feedforward neural network was used to predict AGB from the Barlow Twins-derived representations (NNregressor). To calibrate the NNregressor, the RH-derived footprint-level AGBs from 2019 were used as reference (AGB*RH), and all 32 Barlow Twins representations from the same year were used as predictor variables. Please note: The AGB modeling was done for the following tiles together: 22MCT, 22MCU, 22MDT, 22MDU, 22MGB, 22MGC, 22MHB, 22MHC.For more details, the reader is referred to Atzberger et al. (2025, A scalable, annual aboveground biomass product for monitoring carbon impacts of ecosystem restoration projects, Remote Sensing of Environment). Years 2013-2016 are based on Landsat, 2017-2024 on Sentinel-2 Naming: S2tile_YEARModelApplied_BT_YearModelTrained_GP_YearGEDIPoints.tif e.g. 22MDU_2014_BT_2019_GP_2019.tif