Improving plant-available water estimation using model averaging of national soil water models

Introduction Multiple operational soil water balance (SWB) models provide real-time estimates of soil moisture across Australia, yet differences in model structure and outputs introduce uncertainty for end users. Model averaging offers a potential pathway to improve predictions, but previous studies have largely applied static weighting schemes. This study investigates a temporally dynamic implementation of the Granger–Ramanathan (GRA) model averaging approach to improve in situ and spatial estimates of plant-available water (PAW) in southeastern and southern Australia. Methods Two hypotheses were tested: (1) that GRA model averaging improves point-scale PAW predictions compared to individual models, and (2) that spatially scaling GRA coefficients produces more accurate PAW maps than equal-weight averaging. Soil moisture sensor networks from three study regions were used to evaluate GRA performance at the probe scale. Spatial implementations of GRA were developed using temporally varying coefficients, with and without environmental covariates, and compared against static models and simple averaging. Results At the point scale, GRA consistently outperformed individual SWB models and equal weighting, achieving higher concordance with sensor observations (e.g., mean concordance of 0.87 at Boorowa, 0.73 at Muttama, and 0.90 at Eyre Peninsula, compared to 0.29–0.53 for individual models and 0.05–0.60 for equal weighting). Spatial GRA with dynamic coefficients improved mapping performance relative to static approaches, but incorporating environmental covariates did not consistently enhance accuracy and in some cases reduced model generalizability. Discussion Dynamic GRA model averaging provides a practical framework for integrating multiple national-scale SWB models to improve real-time PAW prediction, particularly at well-instrumented locations. However, scaling these benefits to landscape mapping remains challenging when sensor networks are sparse or unevenly distributed. The approach has potential applications in agricultural decision-making and environmental monitoring, but further refinement is needed to optimise spatial implementations.