A multi-task learning model for global soil moisture prediction based on adaptive weight allocation

期刊论文
出版模式：

开放获取
发表日期：

2025年5月28日

查看全文

Yeguang li,
Haiying Liu,
Tianshan Lv

摘要

Global soil moisture is crucial for multiple disciplines such as earth science and agricultural science. Therefore, there are many studies on how to improve the prediction accuracy of soil moisture, especially the rapid development of deep learning in recent years, which has greatly improved data-driven models. This paper proposes an adaptive weight long short-term memory (AW-LSTM) model based on dynamic weight allocation, which dynamically optimizes the model by calculating the correlation coefficient \(p\) between tasks. The experimental verification demonstrated that the AW-LSTM model exhibited the most accurate soil moisture prediction, with \(R\) and \(R^2\) values of 0.9456 and 0.8489, respectively. These values were 0.019 and 0.072 higher than the single-task prediction values observed in the benchmark dataset.