EGU25-20070, updated on 15 Mar 2025
https://doi.org/10.5194/egusphere-egu25-20070
EGU General Assembly 2025
© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.
Poster | Monday, 28 Apr, 14:00–15:45 (CEST), Display time Monday, 28 Apr, 08:30–18:00
 
vPoster spot A, vPA.19
Enhancing Soil Moisture Estimation through Machine Learning Models and Remote Sensing Data
Vidushi Sharma, Siddik Barbhuiya, and Vivek Gupta
Vidushi Sharma et al.
  • Indian Institute of Technology, Mandi, Indian Institute of Technology, Mandi, School of Civil and Environmental Engineering, India (vidushisharma924@gmail.com)

Moisture content available in soil, is a critical parameter for understanding the health of ecosystems, agricultural productivity, and the management of water resources. Soil moisture is an essential component in the growth of vegetation, climate regulation, and the hydrological cycle. The correct estimation of soil moisture is very crucial for optimizing irrigation, enhancing crop yields, and managing water resources. Spatial coverage limits traditional in-situ measurements, while remote sensing-based approaches, especially using SAR imagery, provide scalability to large-scale spatial coverage for soil moisture estimation. This study compares five machine learning-based models- Long Short-Term Memory (LSTM), Random Forest (RF), Multiple Linear Regression (MLR), Multi-layer Perceptron (MLP), and Support Vector Machines (SVM)-for deriving estimates of soil moisture using features based on VV and VH polarizations and incidence angle from SAR imagery. Model performance was also evaluated using in-situ measurements from Vaira Ranch in California's Central Valley, which comprises grasslands and wetlands. Meteorological data, which include precipitation and antecedent rainfall from the ERA5, were used to improve prediction. Each model was hyperparameter tuned, with LSTM adjusting layers, units, and learning rate; RF optimizing tree number, depth, and feature selection; MLR modifying regularization strength; MLP refining layers, neurons, and activation function; and SVM fine-tuning kernel type, C, and gamma. Performance metrics used for evaluation included R² and Root Mean Square Error (RMSE). The results indicated that LSTM outperformed other models with a R² of 0.89, followed by SVM at a value of 0.81 and RF at a value of 0.78. MLP and MLR values were lower at 0.67. This research focuses on the advantages of the integration of remote sensing data and meteorological information for better soil moisture estimation using machine learning and show that the advanced models such as LSTM and RF can effectively predict soil moisture, with important implications for improving agricultural management and resource planning.

How to cite: Sharma, V., Barbhuiya, S., and Gupta, V.: Enhancing Soil Moisture Estimation through Machine Learning Models and Remote Sensing Data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20070, https://doi.org/10.5194/egusphere-egu25-20070, 2025.