Msc Defense by Walid Ghariani
On Friday, May 26 at 10 a.m. Walid Gahriani will present his Msc Thesis “PM2.5 Prediction Using Remote Sensing and Meteorological Data Across Bavaria” in room 01.027, building 70, Emil-Fischer-Straße 70
From the abstract:
Fine particulate matter (PM2.5) are suspended particles in the air with diameters equal or less than 2.5 μm. Due to its aerodynamic characteristics and small size, PM2.5 can penetrate the human body through the respiratory system causing major human health problems. While PM2.5 monitoring stations provide valuable information with high temporal coverage, most of the monitoring networks are limited in their spatial coverage, limiting the capacity to understand the complete pattern of PM2.5 across a region. This study investigated the potential use of remote sensing and meteorological data coupled with machine learning (ML) for PM2.5 prediction across the state of Bavaria, Germany, at a spatial resolution of 1 km. Two ML models, RF and XGBoost, were developed using training dataset from 2018-2020. For model validation, 10-fold random and spatial cross-validation (CV) were used. The results showed that XGBoost outperformed RF with each CV strategy. With the random CV, XGBoost obtained a moderate performance accuracy with an R² of 0.56, however with the spatial CV, XGBoost prediction decreased at below average performance with an R² of 0.49. For model testing, the dataset from 2021 was used to assess the predictive capability of the models on a daily, weekly and monthly scales. The findings showed that XGBoost outperformed RF at each temporal scale, and the model’s performance improved significantly from a below average performance at a daily scale (R² = 0.47, RMSE = 3.63 μg/m³, MAE = 2.85 μg/m³) to a moderate performance at a weekly scale (R² = 0.64, RMSE = 2.05 μg/m³ and MAE = 1.69 μg/m³) to a strong performance at a monthly scale (R² = 0.71, RMSE = 1.29 μg/m³ and MAE = 1.04 μg/m³). These findings indicate that daily predictions made by XGBoost can be used to produce more accurate predictions over longer temporal scales. A closer look at the seasonal variation of PM2.5 revealed a seasonal pattern with an increase during colder seasons, in autumn and especially during winter, and a decrease during warmer seasons, in spring and during summer.
1st supervisor: Dr. Insa Otto
2nd supervisor: Dr. Michael Thiel
