There is one question that has been at the heart of the climate debate for decades: what actually determines the temperature in a tropical region? Are atmospheric gases the main culprits, or are there other factors that carry greater weight? Safarana Baksh, recognised in the Exact and Natural Sciences category at the EDUCA EDTECH Foundation’s 2026 Academic Excellence Awards, tackled this question using real-world data, advanced models and a methodological approach that combines the best of classical statistics and machine learning.
The starting point
Guyana is a tropical country situated in the north-east of South America, with a complex climate that has been little studied compared with other regions of the world. Its monthly temperature is influenced by multiple factors that interact in ways that are not always obvious: relative humidity, rainfall, the cycles of the ENSO phenomenon, and the presence of atmospheric gases such as carbon monoxide, methane and nitrogen dioxide.
Safarana’s study analysed these relationships over a period of 96 months, from January 2018 to December 2025, integrating temperature data from NASA POWER with atmospheric gas data obtained from the TROPOMI satellite instrument. The result was a high-quality dataset that enabled the application of two predictive models: linear regression and a Random Forest model optimised using Bayesian techniques.
What the data revealed
The results are clear and, in some respects, surprising. Relative humidity proved to be the factor most strongly correlated with temperature, with a coefficient of r = –0.92 — a very strong negative association that reflects the central role played by atmospheric humidity in thermal regulation in the tropics. Precipitation also showed a significant correlation (r = -0.65).
Atmospheric gases, by contrast, played a much more modest role. Carbon monoxide showed a moderate positive correlation (r = 0.68), but the analyses suggest that this relationship reflects shared seasonal factors — such as biomass burning — rather than a direct causal effect. Methane and nitrogen dioxide showed weak and non-significant correlations.
As regards the models, linear regression outperformed the optimised Random Forest on the test data, with an R² of 0.885 compared with 0.830. This result suggests that, once the key climatic variables have been included, the dominant relationships are essentially linear.
The variable importance analysis confirmed that relative humidity accounts for 67.6% of the predictive variability, followed by the ENSO index at 12.4% and precipitation at 10.9%. Atmospheric gases made only a marginal contribution to the models’ performance.
An underlying trend
The temporal analysis revealed clear seasonal patterns, with temperature fluctuations closely linked to wet and dry cycles and to ENSO phases. And one finding that warrants attention: a warming trend of approximately 0.14°C per year over the period studied. This finding should be interpreted with caution given the study’s relatively short time frame, but it points in a worrying direction.
Why it matters
Safarana’s work has implications that extend beyond Guyana. In tropical regions where global climate models have lower resolution and historical data are scarce, integrating satellite observations with exogenous climate variables provides a valuable and replicable methodological framework.
Their findings also challenge a simplified narrative: it is not enough to look at atmospheric gases to understand temperatures in the tropics. Humidity, rainfall and large-scale climate variability are the real key factors. Ignoring them not only weakens the analysis, but can also lead to erroneous conclusions with real-world implications for climate policy.
