Adaptive water body detection: Integrating deep learning, normalised difference water index, and vector data for farm dam water monitoring with OmniWaterMask

Authors

Nicholas J. Wright, Department of Primary Industries and Regional Development, Western AustraliaFollow
John M A Duncan, The University of Western Australia
Nik Callow, University of Western Australia
Sally E. Thompson, The University of Western Australia
Richard J. George, Department of Primary Industries and Regional Development, Western AustraliaFollow

Document Type

Article

Publication Date

9-2025

Journal Title

ISPRS Journal of Photogrammetry and Remote Sensing

ISSN

ISSN: 0924-2716, eISSN: 1872-8235

Keywords

Water detection, Farm dams, Deep learning, Sentinel-2, Landsat, OpenStreetMap

Disciplines

Agricultural Science | Natural Resources and Conservation | Water Resource Management

Abstract

Farm dams are important water security features supporting both agricultural production and the natural environment. In Australia alone, over two million farm dams provide the water resources underpinning rural and regional primary industries with an annual export value of $80 billion. However, monitoring these water bodies to understand water security and vulnerability is challenging, primarily because of their large quantity, size and highly variable spectral signatures. These characteristics result in difficulty determining thresholds for index-based water detection methods and add to the difficulty of creating adequate training datasets for deep learning methods. We present an adaptive approach named OmniWaterMask (OWM) that uses existing mapped water features to optimise the combination of deep learning outputs and a common water index (Normalised Difference Water Index, NDWI) to achieve robust water detection, for both agricultural and other water resources. OWM demonstrates strong performance across multiple datasets and spatial scales, achieving Intersection over Union (IoU) scores of 96.9 % (Sentinel-2), 73.8 % (Landsat) and 90.9 % (National Agriculture Imagery Program, NAIP). When applied to farm dam monitoring in Western Australia using Sentinel-2 imagery, the approach successfully tracks water extent across a range of dam sizes, with Mean Absolute Error (MAE) of 587 m2 when using Sentinel-2 and 785 m2 when using PlanetScope. Our two case studies demonstrate the practicality and scalability of this approach by monitoring water levels in both a single dam and across 7,172 farm dams at monthly intervals over an 8-year period. This methodology enables reliable monitoring of small water bodies at scale, supporting rural water security assessment in increasingly uncertain climatic conditions. The open source OWM library is made available as a Python package on PyPI.

Recommended Citation

Nicholas Wright, John M.A. Duncan, J. Nik Callow, Sally E. Thompson, Richard J. George, Adaptive water body detection: Integrating deep learning, normalised difference water index, and vector data for farm dam water monitoring with OmniWaterMask, ISPRS Journal of Photogrammetry and Remote Sensing, Volume 227, 2025, Pages 714-732 https://doi.org/10.1016/j.isprsjprs.2025.07.007.