Decoding XCO ₂ Distributions Over the Indian Subcontinent Using Deep Neural Network Based High‐Resolution Long‐Term Satellite Observations

Abstract India often faces challenges in monitoring atmospheric carbon dioxide (CO 2 ) through satellite observations due to persistent cloud cover, especially during the monsoon season. This limitation affects the continuous tracking of carbon and hinders accurate assessments of carbon‐climate interactions. To address this, we developed a high‐resolution (0.25°) monthly column‐averaged CO 2 (XCO 2 ) data set for 2003–2020 using a Machine Learning (ML)‐based Deep Neural Network (DNN) downscaling and integration of three satellite retrievals of XCO 2 (SCanning Imaging Absorption SpectroMeter for Atmospheric ChartographY; SCIAMACHY, Greenhouse gases Observing SATellite; GOSAT and Orbiting Carbon Observatory; OCO‐2) across India. The ML‐predicted XCO 2 shows strong agreement with OCO‐2 data for 2018–2020 (correlation coefficient, CC > 0.9; standard deviation: 0.39 ppm), and latitudinal biases range within ±2 ppm. Also, seasonal mean biases are lower compared to global models (CAMS, CT) with values of 0.64 ppm (AMJ), 0.54 ppm (JAS), and −0.03 ppm (ON) except during DJFM (1.06 ppm), suggesting better seasonal consistency. Further, estimated XCO 2 growth rate (GR) from ML (3.73 ppm/year) closely matches NOAA surface observations (3.65 ppm/year) during strong El Niño (2015–2016). Importantly, the inter‐annual variability (IAV) of ML‐XCO 2 GR aligns well with satellite observations (CC = 0.86) outperforming CAMS (0.11) and CT (0.73), indicating the ability of ML in capturing the IAV. The model also accurately detects regional hotspots of CO 2 , especially over the Indo‐Gangetic Plains, consistent with the ODIAC fossil fuel emission inventory. These results demonstrate the ability of ML‐based downscaling for a deeper understanding of regional carbon dynamics and their response to climate variability.