190-11 LEVERAGING ARTIFICIAL INTELLIGENCE IN HYDROLOGY TO PROCESS CITIZEN SCIENCE PHOTOS OF WATER LEVELS

Abstract:

The increasing need for hydrologic data in ungauged watersheds has led to the growth of citizen science initiatives, such as the CrowdHydrology project, where participants submit observations of water levels. While traditionally based on text-messaged numerical data, the widespread use of smartphones has resulted in an influx of unprompted photos of staff gauges, creating an opportunity for image-based processing and quality control. This research develops a methodology to extract water levels and station identification from these images using Large Language Models (LLMs), specifically designed to lower the barrier to entry for professional scientists who may lack expertise in computer vision.

The developed workflow employs a multi-stage process starting with Region of Interest (ROI) extraction using the YOLOv8 object detection model to isolate the staff gauge and the station label. Images are then preprocessed through grayscaling, contrast enhancement, and morphological filtering to reduce noise. A crucial step in the pipeline involves identifying the visual waterline through binarization and scanning, which is then marked with a horizontal red line before being passed to the Gemini 2.5 Flash LLM. The LLM utilizes structured responses via Pydantic data models to return validated numerical values for both water level and station ID.

Results indicate that while "naive" processing of raw images resulted in a Root Mean Squared Error (RMSE) of 0.89 feet, the addition of the visual waterline label significantly improved accuracy, reducing the RMSE to 0.25 feet. Furthermore, the LLM demonstrated a 98% accuracy rate in correctly identifying station IDs. To address potential hallucinations or reasoning errors, the system integrates a human-in-the-loop framework, allowing citizen scientists to verify and update processed contributions. This research demonstrates that leveraging the raw potential of established LLMs can streamline field-based hydrologic observations with minimal technical complexity, though current limitations in image resolution and font size on gauges suggest these methods may be best suited for higher-order streams

Geological Society of America Abstracts with Programs. Vol. 58, No. 5, 2026

doi: 10.1130/abs/2026AM-19464

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.