Detecting Fragmented Flooding in Vegetated Semi-Arid Plains with Satellite Thermal Imagery

Pluvial flooding driven by short-duration, high-intensity rainfall is a dominant hydrological process in semi-arid regions, where surface water is typically sparse and ephemeral. In such environments, flood inundation plays a critical role in regulating surface-groundwater interactions, episodic groundwater recharge, and associated ecological processes. However, these regions are often characterised by a severe lack of ground observation, making remote sensing an essential tool for flood detection and analysis. Flood mapping methods that rely on optical remote sensing are widely used and are well-suited for studying large permanent or semi-permanent water bodies. However, because they rely on threshold-based separation of water from dry land, they perform poorly in detecting shallow, transient inundation, particularly where vegetation obscures the surface water signal.
This study introduces a 1×1 km pixel-based linear interpolation framework for mapping flood dynamics in vegetated semi-arid plains by integrating multi-temporal satellite thermal imagery with topographic constraints. Flood extent is quantified using daily flood-fraction estimates, defined as the proportion of inundated area within each pixel, derived from MODIS diurnal land surface temperature difference. To address data gaps, the framework employs a spatial-temporal reconstruction that combines machine-learning-based interpolation with terrain-informed predictors derived from high-resolution digital elevation models, including elevation metrics, local slope, and the topographic wetness index.
The framework is applied to a terminal basin in the semi-arid region of northwestern Australia, characterised by extensive plains. Rainfall in the region is highly seasonal, occurring predominantly between December and March, with mean daily maximum air temperatures typically ranging from 30 to 40°C. Ten major rainfall-driven flood events since 2002 were analysed. Flood inundation metrics, including inundation duration, event-scale recession characteristics, and cumulative inundation duration, are computed at a resolution of 1 km² to characterise the persistence and spatial connectivity of surface-water bodies.
The results indicate that individual pluvial flood events typically produce inundation lasting approximately 10-20 days, with the decline in total inundated area following an approximately linear recession behaviour. Over the study period (23 yrs), cumulative inundation duration varies spatially across the plain, with individual pixels experiencing between 148 and 166 days of flooding. 
The spatial and temporal variation of flooded area fraction reveals that inundation dynamics across the plain differ fundamentally from the gradual contraction of a continuous lake surface. Instead, the pluvial flood recession appears as a mosaic of short-lived, spatially fragmented water bodies confined to disconnected topographic depressions. Building on this inundation behaviour and integrating flood fraction dynamics with DEM-derived terrain attributes, we develop a predictive model for the spatial distribution of pluvial flooding across the plain.
This study provides a transferable framework for enhancing flood representation and hydrological assessment in data-sparse, dryland environments. By delivering pluvial inundation extents and dynamics, this approach enables improved parameterisation and evaluation of surface hydrodynamic models and surface-groundwater interaction models.