Combining Deep Learning with Physical Parameters in POC and PIC Inversion from Spaceborne Lidar CALIOP
Zhenhua Zhang,Siqi Zhang,Michael J. Behrenfeld,Peng Chen,Cedric Jamet,Paolo Di Girolamo,Davide Dionisi,Yongxiang Hu,Xiaomei Lu,Yuliang Pan,Minzhe Luo,Haiqing Huang,Delu Pan
DOI: https://doi.org/10.1016/j.isprsjprs.2024.05.007
IF: 12.7
2024-01-01
ISPRS Journal of Photogrammetry and Remote Sensing
Abstract:POC and PIC are indispensable components in the global ocean carbon cycle, their transport and space distribution being driven by the biological carbon pump and the carbonate pump. However, passive ocean color remote sensing, usually employed for POC and PIC research, experiences serious shortcomings in polar winter conditions due to its reliance on sunlight, leading to scarce data coverage in the polar regions. In contrast, CALIOP has shown considerable promise in high-latitude ocean observing. Past approaches to estimate POC from CALIOP data relied on bbp measurements obtained through the application of algorithms that presume an empirical linear correlation between bbp and the backscatter coefficient measured at 180°. This method does not account for any spatiotemporal variability in the conversion coefficient. Furthermore, the potential of CALIOP to provide estimates of PIC has not been explored yet. Here, we developed an innovative Two-Branch-Two-Step (TBTS) model to estimate POC and PIC from CALIOP data, which effectively expands the spatial coverage of the MODIS products. This method exploits the strength of deep learning while encapsulating the generalizability of physical parameters. This method consists of two branches: (1) a deep learning branch based on lidar attenuated backscatter waveform and (2) a branch focusing on physical parameters, including the total column-integrated depolarization ratio and the subsurface cross-polarized component of column-integrated backscatter. The model’s generalizability and accuracy are confirmed through the evaluation of a test dataset and validation using in-situ measurements. Our model outperforms several other prevalent machine learning models. We also dissected the importance of different parts of the input data using SHAP tools, thereby providing insights into the black-box nature of deep learning models. Using CALIOP products, we put forth the inaugural estimation of interannually resolved PIC and POC standing stocks in polar regions. The implementation of CALIOP in polar regions bridges the gap inherent in passive ocean color measurements. Lidar-derived total global PIC standing stock is estimated to be 8% higher than that from MODIS, while the POC standing stock is boosted by 17.2%. The carbon standing stock in polar regions exhibits significant inter-annual variability and apparent seasonal periodicity. Hence, results from this research effort clearly reveal that the exploitation of the CALIOP-derived POC and PIC measurements, in combination with the application of new approaches and algorithms to future space lidar data, will undeniably enhance our comprehension of the polar ocean carbon cycle. However, it is important to acknowledge that the CALIOP products may inherit biases from the MODIS data used for training. Hence, where MODIS data is available, it’s still the “better” product to use, but where there isn’t MODIS data, the CALIOP product is extremely useful, especially in the polar regions.