Abstract
Combating climate change has emerged as a global concern recently, and meteorological data remain an important measure for analyzing and predicting climate trends. However, ground weather stations and sensors can be impacted by faults due to accidents and unreliability, often resulting in missing data and lowering the overall quality of the data. This paper explores the impact of using satellite data as an input feature for machine learning algorithms. In particular, temperature, pressure, wind speed, and global horizontal radiation data are imputed using various machine learning algorithms to overcome potential data quality issues resulting from the ground stations. The results from two experiments highlight that the performance of the algorithms significantly increases by using satellite data as input features. For instance, the incorporation of satellite data improved the $R^2$ values for temperature prediction using Random Forest and XGBoost to 0.86 and 0.84, respectively, demonstrating a notable enhancement compared to models without satellite data.
Introduction
Accurate meteorological data is the bedrock of climate change analysis and renewable energy planning. However, ground-level weather stations are notoriously susceptible to failure due to power outages, sensor degradation, or physical damage during extreme weather events. This leads to gaps in the data (missing values) or "noisy" data (inaccurate readings). While statistical methods like ARIMA have traditionally been used to fill these gaps, they often struggle with the complex, non-linear nature of weather patterns. This study investigates whether the consistent, global coverage of satellite data can guide machine learning models in reconstructing missing ground-level measurements.
Methodology
The research framed the imputation problem as a supervised regression task, comparing four machine learning algorithms: K-Nearest Neighbors (K-NN), Random Forest (RF), XGBoost, and Multilayer Perceptron (MLP/ANN).
The study utilized two primary data sources for the year 2014, centered on Toronto International Airport. Ground data was sourced from the Canadian Weather Energy and Engineering Datasets (CWEEDS), providing hourly measurements of temperature, pressure, wind speed, and solar radiation. Satellite data was obtained from NASA’s Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2).
As illustrated in the methodology workflow, the data preprocessing phase involved converting weather data formats, standardizing measurement units, and merging datasets based on timestamps. Two experimental setups were evaluated: a baseline using only cyclic temporal features (hour, day, month) and a satellite-enhanced setup incorporating MERRA-2 measurements as additional input features.
Results and Discussion
The inclusion of satellite data led to a notable enhancement across all meteorological variables. Temperature and solar radiation showed the most dramatic improvements. For temperature, the $R^2$ for Random Forest increased from 0.35 to 0.86. Solar radiation models also performed well, with Random Forest reaching an $R^2$ of 0.80 using satellite inputs.
Visual analysis of the irradiation predictions demonstrates that models incorporating satellite data track the actual irradiation much more closely than baseline models. While baseline models often fail to capture the rapid fluctuations in solar intensity, the satellite-enhanced models leverage the MERRA-2 data to provide a much more accurate reconstruction. While pressure and wind speed remained more challenging due to hyper-local ground conditions, satellite data still significantly improved model utility. Tree-based ensemble models (RF and XGBoost) consistently outperformed the MLP, suggesting that boosting and bagging are more robust for this type of tabular time-series data.
Conclusion
Satellite observations serve as an effective "prior" for machine learning models, capturing trends and correlations that ground sensors might miss during outages. This study proves that integrating satellite data significantly improves the accuracy of climate data imputation. Future research will explore hyperparameter tuning, spatial context from neighboring stations, and advanced architectures like LSTMs to further refine predictive accuracy.
Authors: Kadhim Hayawi, Sakib Shahriar, Hakim Hacid.
Summary updated