Citizen science-Large Language Model alliance for automatic assessment of river hydromorphology

Hydromorphological assessment is a fundamental component of river ecosystem evaluation and management, yet standardized field protocols such as the River Habitat Survey (RHS) require trained personnel and substantial field effort. This study investigates whether multimodal large language models (LLMs) can support preliminary hydromorphological screening based solely on photographic documentation of river reaches originating from citizen science datasets. A dataset of 60 river sections previously assessed using the RHS field protocol was used as a reference. For each reach, photographic images sourced from online mapping services were provided to an LLM, which generated expected values for Habitat Quality Assessment (HQA) and Habitat Modification Score (HMS) through structured reasoning aligned with RHS logic. LLM-derived results were compared with field-based assessments using correlation analysis, agreement metrics, and error statistics. The results indicate moderate agreement between LLM-derived and field-based indices, with the model successfully capturing broad gradients of hydromorphological status. The mean root mean squared error (RMSE) and mean absolute error (MAE) for the HQA assessment were 16.73 and 12.52 points, respectively. For the HMS assessment, the mean RMSE was 28.23, while the MAE was 24.17. For the assessment of the hydromorphological status class of the river, the RMSE was 1.54, and the MAE was 1.28. For the individual assessments of these parameters across different types of photographic exposure used to evaluate the hydromorphological condition of rivers, the highest agreement between the LLM-based assessment and the reference assessment for the HQA parameter and the hydromorphological status class was obtained when photographs taken from high observation points (type D) were used (for HQA: RMSE = 12.95, MAE = 10.30; for hydromorphological status class: RMSE = 0.87, MAE = 0.65). For the HMS assessment, the lowest estimation errors were obtained for exposure types B and C (RMSE = 25.77, MAE = 21.91). Thus, the applied LLM showed the highest agreement in the assessment of the hydromorphological status class and HQA, and the lowest agreement for HMS. HQA estimates showed negligible overall bias, while HMS values tended to underestimate anthropogenic modification. Several RHS parameters related to submerged channel characteristics could not be reliably inferred from photographs, highlighting inherent limitations of image-based assessment. These findings demonstrate that LLM-assisted interpretation cannot replace standardized field surveys for regulatory assessment. However, the approach shows potential as a rapid screening and decision-support tool that may assist river managers in prioritizing sites for detailed investigation and support participatory monitoring initiatives.