Enhancing inundation mapping with geomorphological segmentation: Filling in gaps in spectral observations

Mapping inundation related to environmental water requirements is crucial for the management of Australia's river systems. Remote sensing offers key opportunities for large scale monitoring, but observation of inundation in vegetated areas is challenging due to the vegetation obstructing multispectral sensors on satellites. Several approaches have recently attempted to address this challenge. This study demonstrates a further solution based on extending surface water extent (SWE, inundation) using a novel algorithm. This algorithm uses an initial water mask (usually derived from spectral water classification) and a Probability of Depression Map derived from a LIDAR 10 m and 1 m Digital Elevation Model to extend Sentinel-2 remote sensing derived observations. Using a Probability of Depression Adaptive Threshold (PDAT), the algorithm fills gaps in the remote sensing data by growing initial remote sensing observations (seeds). The algorithm can be easily integrated into a workflow manager for full automation, using a fixed quantile probability for the seed region's PDAT. The method was tested in two contrasting study areas, the Goulburn River in Victoria and the Normanby river in Queensland (Australia), achieving precision and recall values exceeding 80 %. This compares favorably to other algorithms. The results are sensitive to morphological characteristics and elevation variation, performing better for sharply delineated water bodies. Compared to published methods, the geomorphological segmentation algorithm accommodates different water indices and vegetation types. Further development can include automatic selection of an optimal quantile probability based on zonal topographical data that could remove operator bias and further enhance adaptability and accuracy.