Influence of seasonal canopy conditions on ICESat-2-based aboveground biomass estimation in deciduous forests

Forest aboveground biomass (AGB) is a key indicator of terrestrial carbon storage, and its accurate estimation is essential for effective carbon stock monitoring. However, how seasonal canopy conditions (leaf-on vs. leaf-off) affect the accuracy of ICESat-2-based AGB estimation in deciduous forests remains underexplored. This study investigates the effects of seasonal canopy variability on ICESat-2-based AGB estimation and mapping. Predictive models were developed for leaf-on and leaf-off conditions by integrating ICESat-2 photon returns with airborne LiDAR-derived reference AGB, using multiple linear stepwise regression (MLSR) and random forest (RF). Subsequently, 30 m resolution AGB maps were generated by fusing ICESat-2 data with Sentinel-2 and ancillary remote sensing variables using RF and convolutional neural network (CNN) approaches. We further assessed the sensitivity of both estimation and mapping performance to seasonal canopy conditions. Results show that MLSR models produced low estimation accuracy and poor cross-seasonal transferability. RF substantially improved estimation performance compared to MLSR, with R 2 values of 0.53 (leaf-on) and 0.57 (leaf-off), and RMSEs of 37.06 Mg/ha and 34.03 Mg/ha, respectively. The RF model demonstrated moderate cross-seasonal transferability, achieving an R 2 of 0.51 when trained on leaf-off data. CNN further outperformed RF in AGB mapping, increasing R 2 from 0.28 to 0.42 (leaf-on) and from 0.46 to 0.56 (leaf-off), while reducing RMSE to 34.86 and 30.04 Mg/ha, respectively. The leaf-off CNN model achieved the highest accuracy (R 2 = 0.56, RMSE = 30.04 Mg/ha, rRMSE = 21.46 %). These results demonstrate that leaf-off ICESat-2 observations, combined with deep learning approaches, provide clear advantages for improving large-scale forest AGB estimation. • First to assess leaf-on vs. leaf-off effects on AGB estimation using ICESat-2 data • Built AGB models using MLSR and RF under different seasonal canopy conditions • Produced 30 m AGB maps by fusing multi-source remote sensing data using RF and CNN • Leaf-off ICESat-2 data improves AGB estimation accuracy and model transferability. • CNN-based mapping achieves higher accuracy than the RF model.