Land Snow Surface Temperature Estimation Using an Optimized Near Real-Time Passive Microwave Algorithm

The land surface temperature (LST) is a crucial geophysical parameter for understanding cryospheric processes such as snow accumulation, freeze-thaw cycles, and the energy budget. However, most existing passive microwave (PMW) LST retrieval algorithms are not optimized for the land snow surface temperature (LSST) estimation, presenting a significant limitation for remote sensing observation of the cryosphere. To address this, this study developed a robust, all-weather, near real-time, standalone PMW-based LSST retrieval algorithm optimized for snow-covered conditions in the Northern Hemisphere. In this study, LSST is defined as the air temperature close to the snow surface. Due to limited in situ observations in Arctic regions and complex microwave radiative transfer over snow-covered landscapes, the multilayer perceptron (MLP) model, which is referred to in this study as the MLP_Model, was developed. Using 2-m air temperature as a proxy to present the air temperature close to the snow surface, this study employed multitask learning (MTL) to integrate data from in situ automatic weather stations (AWSs), the ECMWF Reanalysis v5 (ERA5) dataset, and simulations from the Microwave Emission Model of Layered Snowpacks (MEMLS) for model training. This integration method balanced the information from multiple data sources, thereby mitigating potential uncertainties associated with training empirical models on limited, single-source datasets while ensuring that the trained model remains broadly consistent with the currently available data and established physical principles. Data from MEMLS simulations act as a physical constraint in the training process, ensuring that the model’s estimates adhere to physical model expectations. The MLP_Model-estimated LSST was compared with LSST from the AWS network and the ERA5 in the Northern Hemisphere for evaluation. The mean absolute error (MAE) values were $3.74~^{\circ }$ C and $4.38~^{\circ }$ C, while 75th percentile absolute errors (Q3AEs) were $5.61~^{\circ }$ C and $5.71~^{\circ }$ C, respectively. Compared with the PMW LSST estimation algorithm developed by Kelly 2003 (Kelly_2003), which has been used in the Japan Aerospace Exploration Agency (JAXA) operational snow retrieval algorithm, the MLP_Model demonstrated a reduction in the MAE by $1.5~^{\circ }$ C and the Q3AE by $2.1~^{\circ }$ C. These results indicate that this newly developed model can provide daily, reliable, and consistent near-real-time LSST estimates based solely on PMW remote sensing observations. The model-estimated LSST could serve as a reliable and independent reference for cryospheric climate studies, offering valuable input for data assimilation and reanalysis efforts. Furthermore, it could support other near real-time operational PMW-based retrievals of cryospheric geophysical parameters such as snow depth (SD) or freeze-thaw states.