Artificial Intelligence Applications in Earth System Modeling and Geoscience: A Hybrid Deep Learning Framework for Precipitation and Soil Moisture Prediction
DOI:
https://doi.org/10.59075/tamsaal.v4i11.280Keywords:
Earth System Models; Artificial Intelligence; Deep Learning; Hybrid CNN–LSTM; Physics-Informed Machine Learning; Precipitation Forecasting; Root-Zone Soil Moisture; Hydrometeorological ModellingAbstract
Earth System Models (ESMs) are the key to gaining insight into the interactions among different components of the Earth system such as atmosphere, oceans, land surface, and cryosphere. Despite the essence of ESMs, their forecasting ability is limited due to the problems of high computational requirements, unaccounted subgrid processes, and uncertainties regarding the use of traditional parameterization schemes. Advances of artificial intelligence (AI), particularly deep learning, could represent a solution for overcoming traditional barrier to using ESMs for prediction purposes since AI encompasses the ability to learn complicated nonlinear relations between different aspects of the Earth system. The present study introduces a hybrid framework based on the use of Convolutional Neural Network–Long Short-Term Memory combination, coupled with the physics-informed loss function for predicting precipitation and root-zone soil moisture over a medium- to short-range period in a heterogeneous mid-latitudinal area. The elements of the framework include the application of various data sources including satellite precipitation data, soil moisture data, and atmospheric reanalysis data from the period of time between 2015 and 2022. The performance of the model is evaluated against persistence, climatology, and numerical weather prediction (NWP)-based benchmarks and measures such as root-mean-square error (RMSE), Pearson’s correlation coefficient (r), Nash–Sutcliffe efficiency (NSE), and critical success index (CSI). Results show that the proposed framework provides an RMSE reduction of 18-24% for 24-hour precipitation and an increase in NSE of 0.12-0.19 for soil moisture when compared to the applicable models. The performance is improving in particular during convective precipitation occurrence as well as speedy changes in the soil moisture after the rainfall. In addition, the inclusion of physical constraints also improves the water budget consistency, while the statistical accuracy is not affected, and adds minor advantages in the independent verification basin. The results of the analysis indicate a potential usage of physics-based hybrid deep learning systems in this area, since this will make it possible to combine the capacity of data-driven models with its fundamental physical drawbacks.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Zohair Aman Khan, Faizan Ur Rehman Qaiser, Sami Ullah Tahir

This work is licensed under a Creative Commons Attribution 4.0 International License.




