Haiyang et al. (2026) From ungauged to poorly gauged sites: A scalable deep learning framework for reference evapotranspiration forecasting through regional generalization and transfer learning

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-09-25
• Authors: Bai Haiyang, Junzeng Xu, Lin Junxian, Qi Wei, Yang Chongguang, Shengyu Chen, Yawei Li, Qianjing Jiang, Qi Zhiming, Muhammad Arif
• DOI: 10.1016/j.agwat.2026.110810

Research Groups

• The State Key Laboratory of Water Disaster Prevention, Hohai University, Nanjing, China
• Jiangsu Province Engineering Research Center for Agricultural Soil-Water Efficient Utilization, Carbon Sequestration and Emission Reduction, Hohai University, Nanjing, China
• College of Agricultural Science and Engineering, Hohai University, Nanjing, China
• Department of Biosystems Engineering, Zhejiang University, Hangzhou, China
• Department of Bioresource Engineering, McGill University, Sainte-Anne-de-Bellevue, QC, Canada
• Department of Agronomy, The University of Agriculture Peshawar, Pakistan

Short Summary

This study develops a scalable deep learning framework for reference evapotranspiration (ET0) forecasting through regional generalization and transfer learning. The framework adapts to local data availability and improves ET0 forecasting performance at ungauged or poorly gauged sites.

Objective

• To develop four deep learning models site-specifically, namely CNN, BiLSTM, CNN-BiLSTM, and CNN-BiLSTM with multi-head attention (CNN-BiLSTM-MHA) for ET0 forecasting.
• To develop four regional generalized ET0 forecasting models based on four deep learning architectures, and to evaluate their performance at held-out stations treated as ungauged during model development.
• To assess the performance of four regional generalized models adapted via transfer learning using limited site-specific data from poorly gauged sites.

Study Configuration

• Spatial Scale: Regional scale in southeastern China.
• Temporal Scale: Daily time series data from 2000 to 2019.

Methodology and Data

• Models used: CNN, BiLSTM, CNN-BiLSTM, and CNN-BiLSTM-MHA.
• Data sources: Meteorological observations from 15 stations in southeastern China, public weather forecasts with lead times of 1–7 days.

Main Results

• The framework achieved improved ET0 forecasting performance at ungauged or poorly gauged sites through regional generalization and transfer learning.
• The CNN-BiLSTM-MHA model showed the best performance for site-specific forecasting, with an average 1–7-day RMSE of 0.26 mm d−1 across 15 stations.
• Transfer learning improved performance at poorly gauged sites, with a maximum gain of 48.58% in RMSE.

Contributions

• The study established an ET0 forecasting framework that adapts to local data availability and improves ET0 forecasting performance at ungauged or poorly gauged sites.
• The framework can extend irrigation scheduling and agricultural water allocation services to regions with uneven monitoring capacity.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 51979012) and the Jiangsu Province Engineering Research Center for Agricultural Soil-Water Efficient Utilization, Carbon Sequestration and Emission Reduction.

Citation

@article{Haiyang2026From,
  author = {Haiyang, Bai and Xu, Junzeng and Junxian, Lin and Wei, Qi and Chongguang, Yang and Chen, Shengyu and Li, Yawei and Jiang, Qianjing and Zhiming, Qi and Arif, Muhammad},
  title = {From ungauged to poorly gauged sites: A scalable deep learning framework for reference evapotranspiration forecasting through regional generalization and transfer learning},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110810},
  url = {https://doi.org/10.1016/j.agwat.2026.110810}
}