An hourly weather forecasting system for specific coordinates in Bangladesh using MERRA-2 reanalysis and machine learning
Bangladesh is highly exposed to climate risk, and agriculture, water resources and disaster management all depend on weather. This study builds a system that forecasts hourly weather values at specific Bangladeshi coordinates, with temperature at 2 meters (T2M) as the target.
The authors collected about 21 years of hourly data (2000 to 2021) from NASA’s MERRA-2 reanalysis through the NASA POWER website. The variables were specific humidity (QV2M), relative humidity (RH2M), surface pressure (PS), wind speed (WS2M), wind direction (WD2M), dew/frost point, and wet-bulb temperature. They used Dhaka district (90.4149°E, 23.7302°N) as the sample location. Day, month and hour were merged into a single hour-of-year (HY) field. A correlation matrix showed that T2M is positively correlated with QV2M and negatively correlated with surface pressure. The authors also made pair plots and time-series plots.
Six classifiers were trained on 70% of the data and tested on the remaining 30%: Support Vector Machine, Random Forest, Logistic Regression, Decision Tree, K-Nearest Neighbors, and Gaussian Naive Bayes.
Test accuracies were:
- Decision Tree: 92.67%
- K-Nearest Neighbors: 91.77%
- Random Forest: 89.4%
- SVM: 72.5%
- Logistic Regression: 27.0%
- Gaussian Naive Bayes: 17.1%
The authors attribute the weak results of Logistic Regression and Naive Bayes to the data being continuous and the variables being interdependent. They conclude that Decision Tree and KNN work best on MERRA-2 data. They also note that temperature alone is not enough for Bangladesh and suggest extending the work to rainfall, floods, cyclones and drought.
