AI-DRIVEN OBJECT DETECTION FROM REMOTE SENSING DATA AND STRATEGIES FOR OPTIMAL MODEL TRAINING RESOURCE ALLOCATION

Abstract

Remote sensing analysis faces challenges detecting subtle atmospheric and oceanic phenomena in noisy datasets. Current AI approaches require extensive computational resources but suffer from inefficient static management. This thesis addresses three interconnected challenges: specialized atmospheric pattern detection, generalized multi-scale object detection, and intelligent resource management. First Contribution: gWaveNet presents a hybrid deep neural network for atmospheric gravity wave detection in single-channel satellite imagery. The frame- work incorporates custom-designed kernels that capture gravity wave characteristics while filtering noise from city lights, clouds, and instrumental artifacts. Results demonstrate superior performance over conventional kernel-based methods and traditional fast Fourier transform approaches. Second Contribution: YOLO-DCAP (Dilated Convolution and Attention- aided Pooling) introduces a generalized multi-scale object detection framework enhancing YOLOv5 through Multi-scale Dilated Residual Convolution (MDRC) and Attention-aided Spatial Pooling (AaSP). MDRC captures multi-resolution features using dilated convolutions while preserving feature propagation through residual connections. AaSP employs attention mechanisms to focus on relevant spatial regions while suppressing interference, effectively handling occlusion and overlap in single-band remote sensing data. Third Contribution: SLA-MORL presents a multi-objective reinforcement learning framework for intelligent resource management in high-performance computing environments. The framework addresses computational infrastructure challenges through hybrid offline-online actor-critic architecture, integrated SLA- driven adaptive optimization, and intelligent event detection with selective trigger- ing. Ablation studies demonstrate 61.1% average improvement in SLA compliance compared to degraded versions, eliminating SLA violations entirely in 44% of test configurations. These integrated contributions provide a comprehensive remote sensing analysis pipeline from specialized algorithms to computational optimization. Experimental validation demonstrates effective combination of domain-specific pattern detection, generalized multi-scale frameworks, and adaptive resource management. This work establishes foundations for next-generation remote sensing systems operating efficiently at scale while maintaining scientific accuracy.