Evaluating the ability of deep learning on actual daily evapotranspiration estimation over the heterogeneous surfaces

Jiaojiao Feng, Weizhen Wang, Feinan Xu and Shengtang Wang

Agricultural Water Management, 2024, vol. 291, issue C

Abstract: Evapotranspiration (ET) is one of the most critical components in hydrological processes and is of great importance to water resource management. Data-oriented deep learning (DL) has been increasingly utilized to forecast hydrological variables over recent years. In this study, the ability of the widely-used DL methods, including long short-term memory (LSTM), bi-directional LSTM (BiLSTM), deep neural network (DNN), and deep belief network (DBN), on the estimation of actual daily ET over the heterogeneous surfaces, was investigated using four groups of experiments. Firstly, the influence of the different input variables on the above DL-based ET model over the various land cover types was explored and analyzed. The results showed that the DL-based ET model can accurately estimate the actual daily ET over the heterogeneous surfaces using a few key conventional observations, i.e., net radiation (Rn), relative humidity (RH), air temperature (Ta), and wind speed (u), as well as soil water content (SWC). However, the performance of the DL-based ET model varied from the different combinations of input variables. SWC was crucial to the estimation of ET, with RMSD decreased from 0.94 to 0.76 mm d−1 when SWC was added to the DL model. Then, the comparison of the four DL-based ET model performances was done. The minor difference in ET estimates was caused by the algorithm differences between LSTM, BiLSTM, DNN, and DBN. Based on the above work, a unified DL-based ET model over the heterogeneous surfaces was developed. The unified DL-based model improved the applicability of DL in ET estimation although it underperformed the separate DL-based model. Finally, the comparison of the performance of the TSEB model and DL method was done. Evaluated with the results of heterogeneous surface, the DL-based model had a better accuracy with a MAPE of 16 – 50%; while the TSEB model had a larger MAPE of 20 – 55%. The results of this research suggest that the DL-based ET model is a promising alternative for the simulation of the daily ET over the heterogeneous surfaces.

Keywords: Data-oriented deep learning; Conventional observations; Soil water content; Comparison of algorithm; Unified DL-based model; Heihe River Basin (search for similar items in EconPapers)
Date: 2024
References: View references in EconPapers View complete reference list from CitEc
Citations: View citations in EconPapers (1)

Downloads: (external link)
http://www.sciencedirect.com/science/article/pii/S0378377423004924
Full text for ScienceDirect subscribers only

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:eee:agiwat:v:291:y:2024:i:c:s0378377423004924

DOI: 10.1016/j.agwat.2023.108627

Access Statistics for this article

Agricultural Water Management is currently edited by B.E. Clothier, W. Dierickx, J. Oster and D. Wichelns

More articles in Agricultural Water Management from Elsevier
Bibliographic data for series maintained by Catherine Liu ().