Machine Learning & Big Data Analyses for Wildfire & Air Pollution Incorporating GIS & Google Earth Engine

• Main Author: Saim, Abdullah Al
• Format: Text
• Language: unknown
• Published: ScholarWorks@UARK 2021
• Subjects: Aerosol Optical Depth; Air Pollution; Arkansas; Fire susceptibility mapping; Google Earth Engine; Machine Learning; Geographic Information Sciences; Nature and Society Relations; Physical and Environmental Geography; Remote Sensing; Aerosol Robotic Network
• Online Access: https://scholarworks.uark.edu/etd/4226; https://scholarworks.uark.edu/cgi/viewcontent.cgi?article=5776&context=etd

The climatic condition, the vegetation type, and the landscape of the United States have made it susceptible to wildfires. This research is divided into two parts based on the analysis of two different aspects of wildfires of two distinct regions. The first part of the study investigates the wildfire susceptibility in Arkansas. Arkansas is a natural state, and it is heavily dependent on its forest and agricultural resources. During the last 30 years, more than 1,000 wildfires occurred in Arkansas and caused more than 10,000 acres of burned areas. Therefore, identifying wildfire-susceptible areas is crucial for ensuring sustainable forest and agricultural resources. Geographic Information System (GIS)-based Machine Learning (ML) can effectively identify fire-prone areas. In this research portion, Multiple Linear Regression (MLR) and Random Forest (RF) methods are applied to 15 layers of GIS data representing natural and anthropogenic factors that influence wildfires. These 15 variables are selected based on the relationship between fire density and explanatory variables. After identifying all variables, geospatial data are prepared and incorporated in RF for training and predicting wildfire-susceptible areas in Arkansas. The obtained R-squared values from RF are 0.99 for the training regression and 0.92 for the validation. Research outcomes suggest that potential evapotranspiration, soil moisture, Palmer Drought Severity Index, and dry season precipitation are the most contributing factors to wildfires in Arkansas among the 15 considered variables. Outputs also indicate that the Ouachita National Forest and the Ozark Forest have the highest susceptibility to wildfires, the southern part of Arkansas has low-to-moderate fire-susceptibility, and the eastern part of the state has the lowest fire susceptibility. The second part of this research investigates the impact of wildfires on air quality over California, which has been chosen for this analysis because of its extensive history of large and severe wildfires. This portion employs the Google Earth Engine (GEE) platform to navigate its geospatial datasets of Moderate Resolution Imaging Spectroradiometer (MODIS) MYD14A1 V6, MCD19A2 Version 6 level 2, and Sentinel-5 Precursor (Sentinel-5P) to validate fire incidents and determine the effect of wildfires on the atmosphere from 2010 to 2020. MODIS MCD19A2 uses an advanced Multi-angle Implementation of Atmospheric Correction (MAIAC) algorithm to produce 1-km resolution images and retrieves Aerosol Optical Depth (AOD) at 470 nm and 550 nm wavelengths. These retrieved AOD values from MODIS are validated using the ground-based sun photometers Aerosol Robotic Network (AERONET), and the uncertainty is checked using the Mean Absolute Error (MAE), the Relative Mean Bias (RMB), and the Root Mean Square Error (RMSE). Linear regression shows good correlations between AERONET and MODIS. The correlation coefficient and the adjusted R-squared value vary from 0.78 to 0.80 and from 0.60 to 0.65, respectively, for AOD values at 550 nm and 470 nm. Results from Sentinel-5P indicate that the 2020 fire events in California raised the NO2 concentration in its atmosphere. This research can improve understanding of the long-term effects of wildfires on air quality and the predictive methodologies that can be used for preemptive measures.