Presentation Information
[AAS14-P01]1DVAR-Like Machine Learning Techniques for Water Vapor Profiling Using Ground-Based Microwave Radiometers
*Kentaro Araki1, Yuya Takashima2 (1.Meteorological Research Institute, JMA, 2.FURUNO ELECTRIC CO., LTD.)
Keywords:
ground-based microwave radiometer,water vapor profile,machine learning,numerical weather prediction data
Ground-based microwave radiometers (MWRs) are essential tools for estimating atmospheric water vapor profiles. While neural networks (NN) and machine learning (ML) are widely adopted for retrieval due to their simplicity, they have long struggled to resolve fine-scale vertical structures, such as water vapor inversion layers or sharp vertical gradients. Conversely, the one-dimensional variational technique (1DVAR), which integrates numerical weather prediction model outputs with MWR observations, generally outperforms ML methods but requires high computational costs for complex radiative transfer calculations.
In this study, we developed two hybrid machine learning approaches inspired by 1DVAR to achieve high-precision water vapor estimation while maintaining low computational costs. These methods leverage information from the Meso-Scale Model (MSM) of the Japan Meteorological Agency. The first approach, ML-CM (Machine Learning with Correction Method), corrects systematic errors in the MSM. The second approach, ML-MT (Machine Learning with Model Training), directly incorporates MSM forecast values as explanatory variables in the training process.
Validation using approximately seven months of data collected in Tsukuba demonstrated that both proposed methods significantly reduced the Root Mean Square Error (RMSE) at all altitudes compared to conventional ML methods trained on reanalysis data. ML-CM successfully corrected systematic biases in the MSM, while ML-MT showed superior accuracy in the lowest atmospheric layer (below 500 m) compared to both the raw MSM forecasts and ML-CM. Case studies confirmed that the proposed methods could reproduce water vapor inversions and steep vertical gradients, which were previously difficult to capture.
Our analysis revealed that while ML-CM is highly dependent on the MSM and limited in extracting short-term fluctuations, ML-MT utilizes the MSM’s vertical structure as a guide while flexibly reflecting real-time MWR observations. Consequently, ML-MT can provide accurate estimates even when the MSM misrepresents the water vapor structure. Furthermore, ML-MT successfully extracted short-term variations at 1-minute intervals, which cannot be captured by the 3-hourly MSM update frequency. These results suggest that ML-MT is a robust and efficient technique for high-resolution atmospheric monitoring.
In this study, we developed two hybrid machine learning approaches inspired by 1DVAR to achieve high-precision water vapor estimation while maintaining low computational costs. These methods leverage information from the Meso-Scale Model (MSM) of the Japan Meteorological Agency. The first approach, ML-CM (Machine Learning with Correction Method), corrects systematic errors in the MSM. The second approach, ML-MT (Machine Learning with Model Training), directly incorporates MSM forecast values as explanatory variables in the training process.
Validation using approximately seven months of data collected in Tsukuba demonstrated that both proposed methods significantly reduced the Root Mean Square Error (RMSE) at all altitudes compared to conventional ML methods trained on reanalysis data. ML-CM successfully corrected systematic biases in the MSM, while ML-MT showed superior accuracy in the lowest atmospheric layer (below 500 m) compared to both the raw MSM forecasts and ML-CM. Case studies confirmed that the proposed methods could reproduce water vapor inversions and steep vertical gradients, which were previously difficult to capture.
Our analysis revealed that while ML-CM is highly dependent on the MSM and limited in extracting short-term fluctuations, ML-MT utilizes the MSM’s vertical structure as a guide while flexibly reflecting real-time MWR observations. Consequently, ML-MT can provide accurate estimates even when the MSM misrepresents the water vapor structure. Furthermore, ML-MT successfully extracted short-term variations at 1-minute intervals, which cannot be captured by the 3-hourly MSM update frequency. These results suggest that ML-MT is a robust and efficient technique for high-resolution atmospheric monitoring.