Parameter-efficient fine-tuning of SAM for SAR flood mapping: comparative performance and encoder–decoder representation analysis

Synthetic aperture radar (SAR) enables flood mapping under cloud and adverse weather, but segmentation remains challenging across heterogeneous environments. Deep-learning models can improve flood delineation, yet their dependence on task-specific training data limits transfer across regions and events. This motivates adapting pretrained vision foundation models to SAR despite the domain gap from RGB imagery. Among these models, we use the Segment Anything Model (SAM) to compare five parameter-efficient fine-tuning (PEFT) strategies, comprising Adapter, Prefix tuning, Prompt tuning, BitFit and Low-Rank Adaptation (LoRA), under a consistent prompting protocol across four held-out test scenes and repeated training configurations. All PEFT strategies outperform unadapted SAM, with flood-class IoU ranging from 32.26% to 57.94% and F1 scores from 46.72% to 69.17%. A hybrid LoRA-BitFit configuration achieves the highest overall IoU and F1, while LoRA favours recall and BitFit precision. Encoder–decoder representation analysis further shows that LoRA and LoRA-BitFit strengthen flood/non-flood separability in the image encoder, whereas LoRA-BitFit maintains high attention-weight separability across several later mask-decoder stages. These results show that SAR adaptation depends not simply on the number of trainable parameters, but on which model components are modified and how adaptation is distributed across the encoder–decoder architecture.