Supervised and unsupervised ship detection for remotely sensed images

This thesis addresses the problem of automated ship detection in aerial and satellite imagery, a task critical to maritime surveillance, port monitoring, and naval intelligence. Two complementary methodologies are proposed to address this challenge, utilizing both labeled and unlabeled data. The first method introduces a tile-based Global-Local RX (GLRX) detector, which enhances classic RX anomaly detection by combining global statistical context with local spatial awareness through distance-weighted tile fusion. This unsupervised approach allows effective anomaly localization without requiring pixel-level labels. The second methodology employs a supervised deep learning approach, fine-tuning the Segment Anything Model (SAM) using Low-Rank Adaptation (LoRA). The SAM-LoRA framework is optimized for ship segmentation using pixel-level masks. Various LoRA ranks and backbone configurations are tested. While ViT-L yields slightly higher performance, the ViT-B backbone with LoRA rank 256 is selected due to its favorable balance between segmentation accuracy and computational efficiency. Together, these two methods offer a robust pipeline for ship detection under both supervised and unsupervised scenarios.