A Deep Learning Approach For Detecting Earth's Surface Anomalies Using Satellite Imagery

Abrupt and accelerated transformations in the Earth's landmass, whether resulting from natural disasters or human activities, pose significant risks to human well-being and infrastructure. Effective humanitarian aid and disaster relief (HADR) efforts demand expeditious and precise detection of such anomalies. To address these challenges, this work introduces a novel approach for detecting earth surface anomalies using single-view Very High Resolution (VHR) satellite imagery and the advanced U-Net++ convolutional neural network architecture. U-Net++ leverages enhanced image segmentation techniques, capturing both local and global features by extracting critical information from satellite images through a hierarchical and multi-scale framework. The architecture's nested and dense skip connections improve segmentation accuracy by preserving fine details and global context simultaneously. Unlike traditional methods, U-Net++ optimizes the process by combining down sampling and feature refinement in its encoder-decoder structure to deliver precise and pixel-level classification of anomalies. This design accelerates detection and enhances efficiency while minimizing data requirements. A binary classification approach is employed to distinguish anomalies effectively at the image level. Comprehensive evaluation is performed using satellite imagery datasets and real-world case studies to benchmark U-Net++ against existing models, demonstrating its superior accuracy and robustness. The architecture strong generalization capabilities make it a powerful tool for real-time anomaly identification across diverse scenarios, including those with limited annotated datasets. This innovative approach highlights the potential of U-Net++ in advancing anomaly detection for earth surface monitoring and disaster management.