Mapping near-real-time soil moisture dynamics over Tasmania with transfer learning

• Published in: Soil Vol. 11; no. 1; pp. 287 - 307
• Main Authors: Widyastuti, Marliana Tri, Padarian, José, Minasny, Budiman, Webb, Mathew, Taufik, Muh, Kidd, Darren
• Format: Journal Article
• Language: English
• Published: Göttingen Copernicus GmbH 08.04.2025; Copernicus Publications
• Subjects: Accuracy; Agriculture; Algorithms; Calibration; Correlation coefficient; Correlation coefficients; Datasets; Decision making; Deep learning; Environmental monitoring; Land cover; Long short-term memory; Machine learning; Mapping; Meteorological data; Moisture content; Multilayer perceptrons; Real time; Remote sensing; Soil dynamics; Soil layers; Soil moisture; Soil properties; Soil sciences; Spatial discrimination learning; Spatial resolution; Surface layers; Temporal variations; Training; Transfer learning; Water; Australia; Tasmania Australia
• ISSN: 2199-398X; 2199-3971; 2199-398X; 2199-3971
• DOI: 10.5194/soil-11-287-2025

Soil moisture, an essential parameter for hydroclimatic studies, exhibits considerable spatial and temporal variability, which complicates its mapping at high spatiotemporal resolutions. Although current remote sensing products offer global estimates of soil moisture at fine temporal resolutions, they do so at a coarse spatial resolution. Deep learning (DL) techniques have recently been employed to produce high-resolution maps of various soil properties; however, these methods require substantial training data. This study sought to map daily soil moisture across Tasmania, Australia, at an 80 m resolution using a limited set of training data. We assessed three modeling strategies: DL models calibrated using an Australian dataset (51 411 observation points), models calibrated using the Tasmanian dataset (9825 observation points), and a transfer learning technique that transferred information from the Australian models to Tasmania using region-specific data. We also evaluated two DL approaches, i.e., multilayer perceptron (MLP) and long short-term memory (LSTM). The models included the Soil Moisture Active Passive (SMAP) dataset, weather data, an elevation map, land cover, and multilevel soil property maps as inputs to generate soil moisture at the surface (0–30 cm) and subsurface (30–60 cm) layers. Results showed that (1) models calibrated from the Australian dataset performed worse than Tasmanian models regardless of the type of DL approaches; (2) Tasmanian models, calibrated solely using local data, resulted in shortcomings in predicting soil moisture; and (3) transfer learning exhibited remarkable performance improvements (error reductions of up to 45 % and a 50 % increase in correlation) and resolved the drawbacks of the two previous models. The LSTM models with transfer learning had the highest overall performance with an average mean absolute error (MAE) of 0.07 m3 m−3 and a correlation coefficient (r) of 0.77 across stations for the surface layer as well as MAE=0.07m3m-3 and r=0.69 for the subsurface layer. The fine-resolution soil moisture maps captured the detailed landscape variation as well as temporal variation according to four distinct seasons in Tasmania. The models were then applied to generate daily soil moisture maps of Tasmania, integrated into a near-real-time monitoring system to assist agricultural decision-making.