Data-driven techniques for temperature data prediction: big data analytics approach.

Area of Science:

• Meteorology and Climate Science
• Data Science and Machine Learning
• Geospatial Analysis

Background:

• Accurate weather data is crucial for climate change studies and meteorological analysis.
• NASA has developed a global weather prediction model, valuable for data-scarce regions.
• Evaluating NASA's model against ground truth is essential for its application.

Purpose of the Study:

• To assess the prediction accuracy of NASA's temperature data against Nigerian Meteorological Agency (NiMet) measurements.
• To explore the potential of developing an improved forecasting model using NASA data.
• To compare the performance of various machine learning algorithms for temperature prediction.

Main Methods:

• Exploratory data analysis of temperature data from NiMet and NASA.
• Calculation of prediction accuracy metrics (MAE, RMSE, R²) for NASA's temperature data.
• Implementation and comparison of five prediction algorithms: Decision Tree, XGBoost, and three MLP variants (LBFGS, SGD, Adam).

Main Results:

• NASA's temperature predictions showed moderate accuracy, with R² values of 0.710 for maximum and 0.620 for minimum temperatures.
• A strong correlation was observed between NASA and NiMet datasets.
• The MLP LBFGS algorithm demonstrated superior performance, significantly reducing MAE and RMSE compared to other models.

Conclusions:

• NASA's temperature data provides a reliable basis for further analysis and model development.
• The MLP LBFGS model, utilizing NASA data, offers enhanced temperature prediction accuracy for the study area.
• This approach can improve meteorological analysis and climate change extrapolation, especially in regions with limited ground-based weather stations.