DESIGN AND DEVELOPMENT OF AI-BASED FRAMEWORKS FOR ENVIRONMENTAL AND GEOSPATIAL DATA ANALYSIS

Abstract

Wildfires, air pollution, and climate change are interconnected environmental challenges with far-reaching consequences for ecosystems, public health, and global climate stability. Wildfires contribute significantly to atmospheric pollution, releasing large quantities of greenhouse gases and particulate matter, accelerating ozone depletion and global warming. This warming effect causes polar ice to melt, reducing Earth's albedo and creating a feedback loop that further exacerbates climate change. Addressing these issues requires advanced monitoring and predictive systems to mitigate their impact effectively. This thesis presents the design and development of AI-based frameworks for environmental and geospatial data analysis, focusing on wildfire risk detection, air pollution prediction, and sea ice classification. The research integrates state-of-the-art deep learning models and remote sensing data to enhance the accuracy and efficiency of environmental monitoring systems. The study introduces the Swin Transformer and IGNITE-NET models for wildfire risk detection, which leverage dynamic receptive field blocks and channel fusion attention mechanisms to improve predictive accuracy while maintaining computational efficiency. These models demonstrate superior performance in classifying fire risk levels using remote sensing imagery, contributing to proactive wildfire management strategies. In the domain of air pollution prediction, the thesis presents the BREATH-Net model, a hybrid deep learning framework that combines Bi-directional Long Short-Term Memory (BiLSTM) networks with Transformer architectures. Using satellite data, this model accurately forecasts nitrogen dioxide (NO₂) concentrations, offering a robust tool for air quality management and public health interventions. The Arctic-Net model is proposed for sea ice classification, integrating Convolutional Neural Networks (CNNs) with attention mechanisms to efficiently classify sea ice types using Synthetic Aperture Radar (SAR) images. The model outperforms existing methods in accuracy and robustness, providing valuable insights for climate research and maritime navigation. The experimental results across all three domains highlight the superior performance of the proposed models compared to traditional approaches. By combining AI with remote sensing technologies, this research contributes to developing scalable, efficient, and accurate environmental monitoring systems. The findings have significant implications for environmental policy-making, disaster management, and climate change mitigation, demonstrating the transformative potential of AI in addressing complex environmental challenges.