Near-Surface Air Temperature Estimation Based on an Improved Conditional Generative Adversarial Network

Jiaqi Zheng, Xi Wu, Xiaojie Li, Jing PengSensors 2024 · DOI

TL;DR

This paper introduces ARM-cGAN, a novel deep learning model that leverages satellite imagery to accurately estimate near-surface air temperature, overcoming the limitations of sparse ground-based weather stations. By integrating a U-Net backbone with self-attention and residual blocks in the generator, and a multi-scale discriminator, ARM-cGAN achieves superior accuracy and robustness in temperature estimation.

What Problem Does This Paper Address?

Accurate near-surface air temperature data is crucial for various fields, from meteorology and ecology to climate change research. However, traditional methods rely on ground-based meteorological observation stations, which are unevenly distributed, leading to data gaps, especially in remote or geographically challenging areas. This unevenness hinders the acquisition of high-resolution spatial temperature distributions. While satellite remote sensing offers broad coverage, effectively translating its data into precise air temperature estimates remains a challenge. Existing methods, including statistical models and earlier deep learning approaches, have limitations in accuracy, transferability, or handling complex environmental factors.

Methods

The core of this work is an improved Conditional Generative Adversarial Network (cGAN), named ARM-cGAN. Here's a breakdown of its key components:

  • Generator: Built upon a U-Net architecture, the generator is enhanced with:
  • Cascaded Residual Blocks: These deepen the network to better capture complex features and improve the extraction of correspondences between satellite data and temperature.
  • Self-Attention Mechanism: Integrated via attention gates, this allows the model to dynamically weigh the importance of different input features, focusing on critical elements for temperature estimation and suppressing irrelevant information.
  • Discriminator: A multi-level and multi-scale spatial feature fusion discriminator is employed. This design addresses the instability and mode collapse issues common in GAN training by allowing the discriminator to analyze input data at various scales, improving its ability to discern subtle details and global structures.
  • Conditional Input: The model uses Fengyun-4A (FY-4A) satellite cloud imagery (specifically channels 12 and 13, which are sensitive to surface and cloud temperatures) as conditional information to guide the generation of temperature maps.
  • Loss Function: The training incorporates both the standard cGAN adversarial loss and an L1 pixel loss. This combination ensures that generated images are not only realistic but also accurately reflect the mapping relationship with the input data.

Datasets

  • Training and Testing Data:
    • FY-4A Satellite Data: Used as conditional input for the generator. Specifically, data from channels 12 and 13 were utilized, with a spatial resolution of 4 km.
  • ERA5 Reanalysis Data: Provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), this dataset served as the ground truth for near-surface air temperature. It has a spatial resolution of 0.25° × 0.25° and a temporal resolution of 1 hour.
  • Data Preprocessing: Bilinear interpolation was used to unify the spatial resolution of FY-4A and ERA5 data to 0.25° × 0.25°. FY-4A data was aligned to the hourly temporal resolution of ERA5.
  • Dataset Split: The study utilized data from January 2020 to December 2022. A training set of 16,979 data points (from 2020-2021) and a testing set of 3400 data points (randomly sampled from 2022) were created, maintaining an approximate 5:1 ratio.

Key Insights and Ideas

  • Synergy of cGANs and Satellite Data: The paper demonstrates that cGANs, designed for image generation, are highly effective when guided by satellite remote sensing data for meteorological tasks like temperature estimation. This approach leverages the generative power of GANs while addressing their inherent lack of control.
  • Importance of Multi-Scale and Attention Mechanisms: The significant improvements achieved by ARM-cGAN highlight the critical role of multi-scale feature fusion in the discriminator and self-attention mechanisms in the generator for capturing complex spatial and temporal temperature variations.
  • Robustness Across Conditions: The ARM-cGAN model shows remarkable consistency and robustness, maintaining high accuracy across different seasons and geographical locations, even in challenging terrains.
  • Addressing Data Scarcity: By effectively filling data gaps from sparse ground stations, this method offers a path towards more comprehensive and reliable near-surface air temperature monitoring.

Key Results

  • Superior Accuracy: ARM-cGAN achieved a Root Mean Square Error (RMSE) of 1.4815 °C and a Pearson Correlation Coefficient (CC) of 0.9897.
  • Significant Improvements: Compared to the baseline CRN model, ARM-cGAN showed a 68.75% improvement in RMSE and a 10.53% increase in CC.
  • Outperformance: The model outperformed other advanced deep learning models, including Attention U-Net, Pix2pix, TransUNet, and DLPE-MS, in both RMSE and CC metrics.
  • Visual Validation: Spatial distribution maps (Figures 6 and 7) and seasonal temperature comparisons (Figure 8) visually confirm ARM-cGAN's superior ability to capture temperature patterns and its consistency across different regions and times.

Limitations and Future Work

  • Geographic and Data Dependency: The current ARM-cGAN model is primarily trained and validated on FY-4A satellite data. Its applicability might be limited in regions where FY-4A data is unavailable or of different quality.
  • Generalization: Further exploration of the model's generalization capabilities across a wider range of geographic environments and atmospheric conditions is needed.
  • Future Directions:
    • Incorporating More Factors: Including additional meteorological factors like wind speed and direction.
  • Multi-Source Data Fusion: Integrating data from other satellite sources (e.g., MODIS, Sentinel, Landsat) to enhance global applicability.
  • Model Optimization: Developing lightweight versions for real-time estimation in resource-constrained environments.
  • Interpretability: Improving the model's interpretability to gain deeper insights into its decision-making process.
  • Climate Prediction Framework: Building a multi-scale framework that combines global climate forecasting with regional micro-scale analysis.

Summary updated

Abstract

To address the issue of missing near-surface air temperature data caused by the uneven distribution of ground meteorological observation stations, we propose a method for near-surface air temperature estimation based on an improved conditional generative adversarial network (CGAN) framework. Leveraging the all-weather coverage advantage of Fengyun meteorological satellites, Fengyun-4A (FY-4A) satellite remote sensing data are utilized as conditional guiding information for the CGAN, helping to direct and constrain the near-surface air temperature estimation process. In the proposed network model of the method based on the conditional generative adversarial network structure, the generator combining a self-attention mechanism and cascaded residual blocks is designed with U-Net as the backbone, which extracts implicit feature information and suppresses the irrelevant information in the Fengyun satellite data. Furthermore, a discriminator with multi-level and multi-scale spatial feature fusion is constructed to enhance the network's perception of details and the global structure, enabling accurate air temperature estimation. The experimental results demonstrate that, compared with Attention U-Net, Pix2pix, and other deep learning models, the method presents significant improvements of 68.75% and 10.53%, respectively in the root mean square error (RMSE) and Pearson's correlation coefficient (CC). These results indicate the superior performance of the proposed model for near-surface air temperature estimation.
BibTeX
@inproceedings{zheng2024near,
  author = {Jiaqi Zheng and Xi Wu and Xiaojie Li and Jing Peng},
  title = {{Near-Surface Air Temperature Estimation Based on an Improved Conditional Generative Adversarial Network}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24185972},
}
Figure