Enhancing forest aboveground biomass carbon estimation using improved phenological features retrieved by pixel-specific fusion of Landsat and Sentinel-2 time series

• Incorporating phenological features to improve forest AGC estimation model. • Propose a pixel-specific fusion method to improve phenological features retrieval. • Improved phenological features alleviated saturation of optical-based AGC estimation. • Higher AGC was linked to delayed phenology and lower spectral variability from SHAP. Temporal analysis from time-series (TS) observations has emerged as an effective approach for improving forest aboveground biomass carbon (AGC) estimation. However, most existing methods generally rely on primary global temporal features of reflectance TS at the annual scale (e.g., annual fluctuation amplitude), while neglecting finer local temporal features within or between seasons, such as forest phenology, which are directly linked to forest carbon sink processes. This underutilization of temporal information constrains AGC estimation accuracy and remains susceptible to saturation effects. To address this gap, we propose a methodology that leverages fine-scale local temporal features (i.e., phenological features) to achieve accurate and reliable large-scale AGC mapping. First, we employed the eXtreme Gradient Boosting (XGBoost) algorithm to verify the feasibility of the phenological features in AGC estimation. Second, we proposed a pixel-specific calibration method to optimally fuse Landsat and Sentinel-2 TS data, enabling more accurate extraction of phenological features. Third, the relationships between phenological features and AGC were characterized using SHapley Additive exPlanations (SHAP) analysis. Validation revealed that the initial model using primary global temporal features produced R 2 of 0.53/0.56 (by Landsat/Sentinel-2, respectively) and RMSE of 26.56/26.28 Mg C/ha. Incorporating phenology-related features significantly improved accuracy, attaining R 2 of 0.60/0.66 and RMSE of 24.86/22.50 Mg C/ha. Further enhancement was achieved by fusing Landsat and Sentinel-2 data using the proposed pixel-specific calibration method, which yielded an R 2 of 0.75 and an RMSE of 19.40 Mg C/ha. This model also partially mitigated saturation effects within the investigated AGC range, reducing RMSE by 14.52 Mg C/ha in high-AGC forests (95.6–170.0 Mg C/ha) compared with traditional models using only global features derived from single-sensor data. Feature importance ranking and SHAP analysis further confirm the critical role of local features, with the five most influential predictors all being phenology-related and exhibiting biologically meaningful relationships with AGC variation. Specifically, higher AGC is associated with delayed phenological timing and reduced short-term spectral fluctuations in reflectance TS data.