This study addresses the challenge of limited in situ observations for carbon cycle analysis when applying machine learning at regional and global scales. Short time series restrict model performance, motivating the testing of data augmentation methods. Satellite-derived data were used as a test to evaluate these approaches. Monthly values of gross primary production, vegetation biophysical indicators, and environmental parameters were reconstructed using MODIS, VIIRS, and TROPOMI data for 2018–2025 during the growing season (May–September) at the Fonovaya observatory in the Tomsk Region. The study evaluates Gaussian copulas, generative adversarial networks, variational autoencoders, and Bayesian neural networks.
carbon cycle, augmentation, satellite data, neural network
1. Liu S., Zhang Z., Zhou H., Chen X. A generative model-based network framework for ecological data reconstruction // Computers, Materials and Continua. 2025. V. 82(1). P. 929-948. https://doi.org/10.32604/cmc.2024.057319.
2. Arshinov M.Yu., Belan B.D., Davydov D.K., Kozlov A.V., Fofonov A.V. Emissiya i pogloschenie parnikovyh gazov lugovoy ekosistemoy yuzhnoy taygi Zapadnoy Sibiri: ocenka vklada pochvennoy sostavlyayuschey po dannym nablyudeniy 2023 g. // Optika atmosfery i okeana. 2024. T. 37(9). S. 760–772. https://doi.org/10.15372/AOO20240906.
3. Skorohodov A.V., Zhuravleva T.B. Fotosinteticheskaya aktivnost' razlichnyh tipov rastitel'nosti yuga Zapadnoy Sibiri i ee vzaimosvyaz' s parametrami okruzhayuschey sredy po dannym reanaliza i sputnikovyh nablyudeniy // Optika atmosfery i okeana. 2026. T. 39(3). S. 266–276. https://doi.org/10.15372/AOO20260311.
