<p>This study presents a novel deep learning framework for analyzing carbon sequestration dynamics in restored forest landscapes using multispectral remote sensing data. We developed a hierarchical convolutional neural network architecture (ForestCarbon-Net) that integrates time-series Sentinel-2 and Landsat imagery with LiDAR-derived structural metrics to quantify above-ground biomass changes and carbon flux dynamics across multiple spatial and temporal scales. Our approach addresses key Limitations in traditional remote sensing analysis by employing attention mechanisms that can distinguish between natural forest regeneration and anthropogenic restoration interventions. Validation across 27 diverse forest restoration sites spanning tropical, temperate, and boreal biomes demonstrates that our model achieves significantly higher accuracy in carbon stock estimation (RMSE: 2.14 tC/ha, R²: 0.93) compared to conventional machine learning approaches. Furthermore, we introduce a new spectral index (Normalized Biomass Accumulation Index) that demonstrates robust performance across seasonal variations and diverse forest types. Longitudinal analysis over a five-year period reveals that the carbon sequestration potential of restored landscapes follows non-linear trajectories dependent on restoration methodology, prior land use, and local climate conditions. Our findings provide crucial insights for optimizing forest restoration strategies and enhancing the accuracy of carbon accounting in natural climate solutions. This framework enables near real-time monitoring of restoration success with implications for improving the effectiveness of global reforestation initiatives and carbon credit verification systems.</p>