14:45 〜 15:00
[AOS12-04] Machine learning for the reconstruction of atmospheric surface humidity over the ocean employing in-situ relevant meteorological measurements
キーワード：relative humidity, data reconstruction, machine learning, regression, climate, atmosphere
Air humidity in the near-surface layer of the atmosphere over the ocean is a key climate parameter that has a significant impact on the processes of moisture and heat transfer between the ocean and the atmosphere, as well as on the dynamics of atmospheric processes in general. Analysis of meteorological data collected during the 20th century shows the sparseness of humidity measurement series in space and time. The International Ocean and Atmosphere Data Set (ICOADS) indicates an insufficient density of measurements in the early 20th century compared to later periods, which creates difficulties for adequate analysis of climate trends in relative humidity.
Methods for approximating humidity time series presented in the literature often demonstrate limited accuracy, based mainly on statistical and heuristic approaches. Our work is aimed at improving the quality of solving this problem through the use of machine learning methods.
As a first, simplest approach, we solved the problem in the formulation of the approximation of relative humidity based on the data of accompanying measurements of atmospheric pressure, air temperature, wind speed and direction, sea surface temperature, as well as observations of the amount and types of clouds. In addition, the accompanying variables include the WMO weather code and the solar altitude angle. In our study, we employed the following types of machine learning models: linear regression, decision trees, random forests, gradient boosting, and fully connected artificial neural network. To improve the territorial and temporal specificity of the developed models, we conducted a study for each 2x2-degree spatial square and for each season separately. The scikit-learn library and the package implementing the CatBoost model were used to train and apply the machine learning models. For each type of model, we optimized the hyperparameters using the Optuna Bayesian optimization library. Based on the results obtained, maps of the spatial distribution of model errors were constructed, which made it possible to identify regions with high and low accuracy of humidity approximation.
The study confirmed the effectiveness of machine learning methods for reconstructing climate series, identified the most suitable models for this task, and outlined promising areas for further work.
Methods for approximating humidity time series presented in the literature often demonstrate limited accuracy, based mainly on statistical and heuristic approaches. Our work is aimed at improving the quality of solving this problem through the use of machine learning methods.
As a first, simplest approach, we solved the problem in the formulation of the approximation of relative humidity based on the data of accompanying measurements of atmospheric pressure, air temperature, wind speed and direction, sea surface temperature, as well as observations of the amount and types of clouds. In addition, the accompanying variables include the WMO weather code and the solar altitude angle. In our study, we employed the following types of machine learning models: linear regression, decision trees, random forests, gradient boosting, and fully connected artificial neural network. To improve the territorial and temporal specificity of the developed models, we conducted a study for each 2x2-degree spatial square and for each season separately. The scikit-learn library and the package implementing the CatBoost model were used to train and apply the machine learning models. For each type of model, we optimized the hyperparameters using the Optuna Bayesian optimization library. Based on the results obtained, maps of the spatial distribution of model errors were constructed, which made it possible to identify regions with high and low accuracy of humidity approximation.
The study confirmed the effectiveness of machine learning methods for reconstructing climate series, identified the most suitable models for this task, and outlined promising areas for further work.