Convolutional Long Short-Term Memory network for generating 100 m daily near-surface air temperature

Area of Science:

• Urban Climatology
• Environmental Science
• Data Science

Background:

• Global warming and urbanization significantly impact urban climates.
• Fine-scale climate studies are crucial for understanding temperature variations in developed cities.

Purpose of the Study:

• To generate a high spatiotemporal resolution dataset of near-surface air temperatures for urban areas.
• To support urban climate research and inform heatwave mitigation policies.

Main Methods:

• Utilized the Convolutional Long Short-Term Memory (ConvLSTM) deep learning model.
• Integrated multi-source data: ERA5 temperature, topography, landcover, and vegetation fraction.
• Achieved a spatial resolution of 100m for daily temperature data (max, min, mean).

Main Results:

• The developed dataset covers summer months (June-August) from 2019-2023.
• High accuracy was confirmed with MAE (0.564-0.784°C), RMSE (0.733-1.027°C), and R² (0.892-0.943).
• The dataset provides recent daily temperature data at a 100m resolution.

Conclusions:

• The high-resolution dataset is valuable for urban and inner-urban climate studies.
• This resource can inform heatwave mitigation and adaptation strategies in cities.
• The study highlights the utility of deep learning for generating detailed climate data.