Mapping evapotranspiration using satellite and eddy covariance in south-western Australia

Evapotranspiration (ET) is the dominant flux in the terrestrial water balance of Mediterranean-type climates and a primary control on groundwater recharge. In south-western Australia, a long-term decline in winter rainfall combined with increasing evaporative demand and urban growth has intensified pressure on groundwater resources that support endemic ecosystems, irrigated agriculture, and urban water supply. Reliable, spatially distributed ET estimates are therefore crucial for groundwater modelling and water resources management under a changing climate.

Despite the widespread availability of satellite-derived ET products, their direct application in regions dominated by endemic vegetation remains problematic. Banksia woodlands, which cover large parts of the Swan Coastal Plain, exhibit deep rooting systems, strong soil-vegetation feedbacks, and seasonal water use strategies that are poorly represented in global ET algorithm. Consequently, commonly used products such as MOD16 and PML show significant discrepancies in magnitude and seasonal dynamics, leading to large uncertainty in groundwater recharge estimation.

To address this limitation, we developed a locally constrained ET upscaling framework that integrates multiple satellite products with ground-based observations across the Swan Coastal Plain. Empirical regression relationships were first derived for MOD16 and PML ET estimates to characterise systematic product differences. Then, time series were used to train and apply a Random Forest (RF) model, constrained by eddy covariance observations. Finally, in Google Earth Engine (GEE) the ET was upscaled in space and time using satellite-based predictors and land-cover information.

This contribution presents a multi-year, monthly ET climatology for the Perth region and evaluates its spatial and temporal consistency across major land-cover classes, with particular emphasis on banksia woodland ecosystems. Rather than benchmarking individual products alone, we assess the implications of ET uncertainty and upscaling choices for groundwater recharge estimation and regional groundwater modelling.

The resulting ET maps reveal systematic biases in standalone MOD16 and PML products over Banksia woodlands and demonstrate that the RF-based upscaling produces more coherent seasonal patterns and spatial gradients consistent with field observations. In particular, the RF model systematically constrains the high ET values characteristic of PML while preserving the spatial structure captured by MOD16. Monthly mean ET fields show reduced inter-product variability and offer stable behaviour suitable for direct use as inputs to groundwater modelling.

These results indicate that combining satellite-derived ET products through locally informed regression and machine-learning upscaling substantially improves the representation of evapotranspiration in groundwater modelling frameworks. The derived ET climatology provides a defensible basis for recharge estimation and scenario analysis under ongoing and projected climate evolution in south-western Australia. Nevertheless, more eddy covariance sites would improve estimates.

More broadly, this approach offers a transferable framework for adapting global ET products to endemic and water-limited ecosystems, supporting more robust groundwater-resource management in regions facing increasing hydroclimatic stress.