A Comparative Study of Flood Monitoring Techniques Using the UN-SPIDER Recommendations and a Generative AI-Based Model (SATGPT): A Case Study in Ayutthaya, Thailand

Abstract

This study presents a dual-track workflow for systematic flood mapping and analysis in Ayutthaya Province, Thailand, a critical region frequently subjected to severe flooding from monsoonal rainfall and the overflow of the Chao Phraya River. The research detected and analyzed multiple flood events that occurred over a significant five-year period, from 2016 to 2020. The objective was to compare the flood extent derived from the UN-SPIDER recommendations using SAR imagery with that obtained from the generative AI-based model known as SATGP. The first technique utilized a physics-based change detection method applied to Sentinel-1 Synthetic Aperture Radar (SAR) imagery, executed entirely on the Google Earth Engine (GEE) cloud computing platform. To ensure high delineation accuracy for inundated areas, the methodology incorporated rigorous post-processing filters. These steps included: applying a harmonized ratio threshold of 1.25 to the post-flood/pre-flood image ratio; masking out permanent water bodies using the JRC Global Surface Water dataset; and excluding areas with a slope greater than five percent via a Digital Elevation Model (DEM) to minimize topographic errors. The second technique explored an innovative AI-assisted retrieval approach, employing SATGPT. This cutting-edge geospatial decision-support tool leverages generative artificial intelligence to efficiently translate complex natural-language prompts into executable geospatial analyses and final raster outputs. Comparative analysis confirmed that both methods successfully generated detailed flood maps, yet they displayed distinct spatial patterns. The GEE/SAR product typically mapped medium-to-large, spatially continuous patches primarily concentrated within the province's principal flood risk zones. In contrast, the SATGPT-derived maps showed higher fragmentation with dense, fine-scale patches that closely followed the intricate network of canals and field-bunds, resulting in greater pixel-level coverage but less spatial continuity. Spatially, recurrent flood hotspots were consistently identified in the western low-relief floodplain and along the province's northern corridor. This study validates the complementary value of this dual-track methodology, where the GEE/SAR output provides an authoritative baseline for event delineation, while the SATGPT product offers utility for rapid triage, visualization, and actionable stakeholder briefings.