A Comparative Assessment of Threshold-Based and Machine Learning Methods for Flood Detection

 

Flood mapping plays a key role in understanding hazard impacts, supporting emergency response, and guiding long-term risk planning. Remote sensing is now widely used in flood studies because it offers low-cost data, avoids the need for dangerous field surveys, and provides rapid observations over large areas. Despite these advantages, comparative research remains limited, particularly with respect to differences among flood-mapping algorithms, such as machine-learning versus threshold-based approaches, and the performance of optical versus radar sensors. This research addresses these gaps by applying multiple flood-mapping methods to the same flood event in Pakistan, and then comparing their performance with respect to a validation benchmark to provide a clearer insight into how data selection and methodological design influence flood detection outcomes

This study evaluates four distinct methods for mapping floods using multi-sensor satellite data. To ensure a fair comparison, three unsupervised machine-learning approaches including a synergetic Sentinel-1 and Sentinel-2 workflow, a method integrating harmonized Landsat–Sentinel data with radar, and a daily MODIS imagery technique were tested alongside a traditional Otsu thresholding baseline. All four were tested on the same 2025 Pakistan flood event, characterized by intense monsoon rains and flash flooding across regions such as Sindh and Punjab in mid- to late-2025.  The flood maps were then validated against UNOSAT flood reports for this event, where UNOSAT’s flood extent closely matches the results produced by the Sentinel-1/Sentinel-2 workflow, which yields the most conservative flood extent among the tested methods.

 Larger flood extents from some methods, especially the Sentinel-1 Otsu thresholding approach, include areas not clearly flooded in optical images. This happens because SAR backscatter also responds to wet soil and saturated vegetation, which a simple threshold can misclassify as water, leading to flood overestimation.

Overall, the results show that flood maps are not just different versions of the same answer, they reflect different satellite data and the utilized algorithms detect flooding. Approaches that combine multiple data sources with machine-learning strike a better balance, producing flood extents that are both spatially consistent and physically realistic. This indicates that multi-sensor, machine-learning–based methods are better suited for operational flood monitoring than simple thresholding, which is too sensitive to surface noise and often overestimates flooding.