Ma et al. (2025) Estimation of All-Weather Daily Surface Net Radiation over the Tibetan Plateau Using an Optimized CNN Model

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-11-30
• Authors: Bin Ma, Yaoming Ma, Weiqiang Ma
• DOI: 10.3390/rs17233894

Research Groups

Not specified in the provided text.

Short Summary

This study developed and optimized a deep learning framework using 19 CNN architectures for accurate daily surface net radiation (Rn) estimation over the Tibetan Plateau, finding Xception to be the most effective with high accuracy (R² > 0.94) and superior performance compared to existing products.

Objective

• To develop an optimized deep learning framework for accurate daily surface net radiation (Rn) estimation over the complex and heterogeneous terrain of the Tibetan Plateau.

Study Configuration

• Spatial Scale: Regional (Tibetan Plateau) with per-pixel estimation.
• Temporal Scale: Daily.

Methodology and Data

• Models used: Optimized deep learning framework, systematically evaluating 19 Convolutional Neural Network (CNN) architectures, with Xception identified as the best performer. A per-pixel multivariate regression design (1 × 1 × 21) was used, incorporating engineered neighborhood descriptors.
• Data sources: Multi-source inputs, including surface variables and neighborhood descriptors (specific types like satellite, observation, reanalysis are not detailed but implied by "multi-source inputs").

Main Results

• Xception architecture achieved the best performance with high accuracy (coefficient of determination, R² > 0.94), computational efficiency, and physical consistency.
• The framework demonstrated stable performance across diverse weather conditions and significantly outperformed the GLASS product, particularly over rugged terrain and high-albedo surfaces.
• SHAP analysis revealed that astronomical and topographic factors contributed approximately 70% to Rn predictions, while surface properties contributed approximately 25%.
• The depthwise separable convolutions and skip connections of Xception effectively captured complex dependencies between surface variables and Rn through hierarchical nonlinear cross-channel feature learning.

Contributions

• Development of an optimized deep learning framework for daily Rn estimation that systematically evaluates a wide range of CNN architectures.
• Introduction of a per-pixel multivariate regression design with engineered neighborhood descriptors to embed spatial context, avoiding common artifacts of patch-based methods.
• Identification of Xception as a highly accurate, efficient, and physically consistent model for Rn estimation over complex terrain.
• Demonstrated superior performance compared to existing products (e.g., GLASS), especially in challenging environments like rugged terrain and high-albedo surfaces.
• Provided insights into the physical drivers of Rn predictions through SHAP analysis, quantifying the contributions of astronomical, topographic, and surface factors.

Funding

Not specified in the provided text.

Citation

@article{Ma2025Estimation,
  author = {Ma, Bin and Ma, Yaoming and Ma, Weiqiang},
  title = {Estimation of All-Weather Daily Surface Net Radiation over the Tibetan Plateau Using an Optimized CNN Model},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17233894},
  url = {https://doi.org/10.3390/rs17233894}
}