Shatabda, SwakkharAli, Syeda NazifaRimo, Fahmida AkterHalder, Amit Kumer2026-08-062026-08-0620262026-01ID 22101025ID 21301180ID 21201770https://hdl.handle.net/10361/28825This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science and Engineering, 2026.Cataloged from PDF version of thesis.Includes bibliographical references (pages 64-66).Prediction of global ocean–atmosphere CO2 exchange is often limited by sparse in situ observations and the indirect nature of satellite measurements, leading to substantial uncertainty and disagreement among biogeochemical models, observationbased reconstructions, and atmospheric inversion products in both flux magnitude and variability. We develop a hybrid, time-aware machine learning framework that reconstructs monthly global 1◦×1◦ CO2 flux by learning relationships between air–sea exchange and multi-source environmental drivers. The learning target is the Jena CarboScope ocean flux product derived from SOCAT, which provides spatially and temporally continuous flux fields during 2014–2024 in 2◦×2◦ and serves as a benchmark consistent with carbon-budget constraints. Predictors integrate physical ocean state from ORAS5 (sea surface temperature, salinity, mixed-layer depth, sea surface height), atmospheric forcing from ERA5 (10 m zonal and meridional winds), and biogeochemical indicators from climate-quality products (chlorophyll-a, phosphate, and net primary production), harmonized to a common monthly grid with seasonality encoded via cyclic month terms and standardized preprocessing. Prediction is formulated as supervised regression at the grid-cell level and evaluated under strict temporal generalization using expanding leave-one-year-out windows with train/validation/test splits to prevent leakage; raw-driver and anomalyaugmented representations are compared to assess the value of emphasizing interannual departures or physics-informed features. Results show that nonlinear treebased learning generalizes most robustly across unseen years, with Random Forest achieving the lowest mean window errors (actual dataset: MSE 1.067×10−7, RMSE 3.26 × 10−4, MAE 2.16 × 10−4; anomaly-augmented dataset: MSE 9.829 × 10−8, RMSE 3.13×10−4, MAE 2.06×10−4) and best performance in the longest training window (2014–2022 training; 2024 test), indicating improved transfer with increased historical variability. Beyond reconstruction, predicted flux fields are aggregated using RECCAP2 basin masks and conditioned to deseasoned, detrended interannual variability (IAV) signals to examine changes in basin-scale source–sink behavior; regime shifts are detected conservatively using the nonparametric Pettitt changepoint test with 12-month persistence confirmation, yielding three candidate shifts (Atlantic: 2016-06; Southern Ocean: 2020-10; Arctic: 2021-07), and interpreted using Ni˜no3.4 ENSO as a physical coherence check. This hybrid framework provides a reproducible pathway for global CO2 flux monitoring that integrates multi-source predictors with observation-based anchoring, supports climate-relevant variability analysis beyond headline accuracy metrics, and establishes foundations for future early-warning prediction using shift labels and precursor indicators.78 pagesen-USAttribution-NonCommercial-NoDerivatives 4.0 InternationalBRAC University theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission.http://creativecommons.org/licenses/by-nc-nd/4.0/Supervised regressionInterannual variabilityOcean-atmosphere exchangeCO2 exchangeEnvironmental monitoringTime-series analysisOcean-atmosphere interaction.Neural networks (Computer science).Machine learning.Environmental monitoring--Remote sensing.Global environmental change--Remote sensing.Remote sensing--Environmental aspects.Thesis