Enhancing community-based participatory flood imagery using an AI-based super-resolution framework

Imagery and videos contributed by local communities provide valuable ground-level perspectives of disaster conditions, particularly in suburban and rural areas where monitoring infrastructure is sparse. However, community-based participatory visual data are often degraded by low resolution, motion blur, compression artifacts, and inconsistent metadata. These limitations are further compounded when imagery and videos are captured using older or lower-capability mobile devices with limited sensors and aggressive compression settings. Together, these constraints reduce the interpretability and analytical usefulness of participatory disaster imagery. This study investigates whether domain-adaptive super-resolution can improve the perceptual quality of flood-related images and videos collected under real-world conditions. We fine-tuned a pretrained Real-ESRGAN model using a curated subset of participatory flood imagery and evaluated multiple training configurations. Performance was assessed using no-reference perceptual quality metrics, including CLIP-IQA and MUSIQ, alongside qualitative visual comparison. Results demonstrate consistent improvements in texture preservation, boundary clarity, and structural coherence, with lightweight fine-tuning configurations yielding the strongest and most stable performance. Perceptual gains were more pronounced for video frames than for still images, suggesting that repeated structural cues may help stabilize enhancement under degradation. A case study in Fort Hancock, Texas, illustrates how targeted fine-tuning improves the clarity and interpretability of community-based participatory flood imagery collected in data-scarce environments. Super-resolution is treated as a preprocessing step that enhances image quality and supports subsequent analysis, rather than as a standalone decision-making tool. These findings show that super-resolution performs better when it is trained on imagery that reflects the actual visual conditions found in community-collected disaster data. In practical terms, the approach may help review community-submitted flood imagery more clearly and support downstream tasks such as visual flood extent interpretation, road condition review, and archive improvement, but enhanced outputs should be treated as visual aids rather than ground truth.