An Interpretable framework of Soil moisture estimation based on Mixture-of-Experts (ISMoE): a case study on the Tibetan Plateau

Yang, Junran, Yang, Qinli, Hu, Feichi and Shao, Junming 2025. An Interpretable framework of Soil moisture estimation based on Mixture-of-Experts (ISMoE): a case study on the Tibetan Plateau. Journal of Hydrology 661 (Part C) , 133763. 10.1016/j.jhydrol.2025.133763 Item availability restricted.

PDF - Accepted Post-Print Version Restricted to Repository staff only until 23 June 2026 due to copyright restrictions. Download (3MB)

Abstract

Soil moisture plays a critical role in regulating land–atmosphere exchanges, but its accurate estimation still remains challenging. To better understand soil moisture dynamics and estimate soil moisture more accurately, this study proposes an Interpretable framework of Soil moisture estimation based on Mixture-of-Experts (ISMoE) and investigates its performance across three climatic zones on the Tibetan Plateau. The MoE model consists of multiple expert models and a routing network. Each expert model is designed as an independent deep spatial–temporal neural network (CNN-LSTM) that captures the unique contribution of a specific environmental factor to soil moisture dynamics. The routing network then dynamically weighs the expert model outputs to generate a final integrated soil moisture estimation. Results show that ISMoE achieved CC values of 0.858, 0.863, and 0.958 in NGARI, MAQU, and NAQU regions, respectively. ISMoE outperformed multiple baselines (standard CNN-LSTM model, mean_expert ensemble model, SMAPL3, and ERA5-Land data). The model’s interpretability was validated through the routing network’s weights. The key, region-specific drivers assigned with high importance (weight) by the model are as follows: snow and radiation processes in the arid NGARI region; lake thermal effects and precipitation in the tundra NAQU region; and wind patterns and lake mixing processes in the monsoon-influenced MAQU region. The identified key drivers align with previous study, and the weights revealed distinct, temporally consistent seasonal patterns in each region. The proposed ISMoE improved both accuracy and interpretability of soil moisture estimation model, which provides a promising way to enhance model reliability and valuable dataset for regional water resources research.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Earth and Environmental Sciences
Publisher:	Elsevier
ISSN:	0022-1694
Date of First Compliant Deposit:	26 September 2025
Date of Acceptance:	22 June 2025
Last Modified:	29 Sep 2025 09:15
URI:	https://orca.cardiff.ac.uk/id/eprint/181347

Actions (repository staff only)

Edit Item