Spatial context importance in deep learning for global horizontal irradiance estimation from satellite imagery
Abstract
Accurate estimation of Global Horizontal Irradiance (GHI) is essential for solar energy applications, climate modeling, and various geophysical processes. Traditional satellite-based methods rely on the pixel-centric Independent Pixel Approximation (IPA), which treats each pixel as radiatively isolated from its neighbors. This pixel-centric approach presents challenges, which are amplified by the higher spatial resolutions of third-generation geostationary satellites. In this study, we evaluate deep-learning models that explicitly incorporate spatial context through multispectral image patches of varying sizes ( 3 × 3 to 31 × 31 pixels) from GOES-16 imagery to improve satellite-based GHI estimation and address IPA shortcomings. We compare two architectures—a Fully Connected Network (FCN) and a convolutional-based model—against a state-of-the-art physical retrieval method (PSM3), using in-situ GHI measurements from 31 stations across North America. Our results show that spatial-context-aware deep learning models outperform PSM3 across most metrics, especially under cloudy and partially clear conditions, yielding improved performance, stability, and reduced bias. The best-performing model achieves a 14.6% lower root mean squared error and a 17.1% lower mean absolute error compared to PSM3 on a year-long test set composed of 6 test stations. However, our current implementation of deep learning models still struggles to consistently outperform PSM3 bias under clear-sky conditions at certain sites. By quantifying how image size and architecture jointly shape the performance and generalization of our models, this work underscores the potential of spatial-context-aware deep learning for satellite-based all-sky GHI retrieval.
Related
- Horizontal Strain Rate estimation from InSAR Observations and the Strapdown Method
- Joint Estimation of the Global Facility-Scale Oil–Gas Methane Emission Rate Distribution and Survey-Mode Detection Probabilities from Multiple Satellite Systems
- Composable multi-satellite precipitation estimation for evolving observing systems
- A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data
- SATFRONT: Detecting Atmospheric Fronts with Satellite Data using Machine Learning
BibTeX
@article{becquet2026spatial,
title = {{Spatial context importance in deep learning for global horizontal irradiance estimation from satellite imagery}},
author = {Vadim Becquet and Philippe Blanc and Yves-Marie Saint-Drenan and Yehia Eissa and Hadrien Verbois},
journal = {Solar Energy},
year = {2026},
month = sep,
doi = {10.1016/j.solener.2026.114844},
url = {https://minesparis-psl.hal.science/hal-05742494v1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Spatial context importance in deep learning for global horizontal irradiance estimation from satellite imagery
AU - Vadim Becquet
AU - Philippe Blanc
AU - Yves-Marie Saint-Drenan
AU - Yehia Eissa
AU - Hadrien Verbois
PY - 2026
DA - 2026/09/08
JO - Solar Energy
DO - 10.1016/j.solener.2026.114844
UR - https://minesparis-psl.hal.science/hal-05742494v1
AB - Accurate estimation of Global Horizontal Irradiance (GHI) is essential for solar energy applications, climate modeling, and various geophysical processes. Traditional satellite-based methods rely on the pixel-centric Independent Pixel Approximation (IPA), which treats each pixel as radiatively isolated from its neighbors. This pixel-centric approach presents challenges, which are amplified by the higher spatial resolutions of third-generation geostationary satellites. In this study, we evaluate deep-learning models that explicitly incorporate spatial context through multispectral image patches of varying sizes ( 3 × 3 to 31 × 31 pixels) from GOES-16 imagery to improve satellite-based GHI estimation and address IPA shortcomings. We compare two architectures—a Fully Connected Network (FCN) and a convolutional-based model—against a state-of-the-art physical retrieval method (PSM3), using in-situ GHI measurements from 31 stations across North America. Our results show that spatial-context-aware deep learning models outperform PSM3 across most metrics, especially under cloudy and partially clear conditions, yielding improved performance, stability, and reduced bias. The best-performing model achieves a 14.6% lower root mean squared error and a 17.1% lower mean absolute error compared to PSM3 on a year-long test set composed of 6 test stations. However, our current implementation of deep learning models still struggles to consistently outperform PSM3 bias under clear-sky conditions at certain sites. By quantifying how image size and architecture jointly shape the performance and generalization of our models, this work underscores the potential of spatial-context-aware deep learning for satellite-based all-sky GHI retrieval.
SN - 0038-092X
SN - 1471-1257
ER -