Soil Moisture Retrieval in Farmland Areas with Sentinel Multi-Source Data Based on Regression Convolutional Neural...

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Agricultural Science

Background:

• Soil moisture monitoring is crucial for agriculture, impacting crop growth, yield estimation, and irrigation management.
• Sparse vegetation cover in farmlands presents challenges for accurate soil moisture retrieval using remote sensing data.
• Integrating multi-source remote sensing data offers a promising approach to overcome these limitations.

Purpose of the Study:

• To quantitatively retrieve soil moisture content by combining Sentinel-1 radar and Sentinel-2 optical satellite data.
• To mitigate the influence of vegetation cover on soil moisture estimation.
• To evaluate the performance of different machine learning models and the impact of polarimetric decomposition features on retrieval accuracy.

Main Methods:

• Applied the Oh model and water cloud model to remove vegetation influence.
• Utilized Support Vector Regression (SVR) and Generalized Regression Neural Network (GRNN) to model relationships between remote sensing features and soil moisture.
• Developed a Regression Convolutional Neural Network (CNNR) model, incorporating polarimetric decomposition features from Sentinel-1 PolSAR data, for enhanced deep-level feature extraction.

Main Results:

• SVR achieved an R² of 0.7619 and RMSE of 0.0257 cm³/cm³.
• GRNN achieved an R² of 0.7098 and RMSE of 0.0264 cm³/cm³.
• The CNNR model with optimal features demonstrated superior performance, reaching an R² of 0.8947 and RMSE of 0.0208 cm³/cm³. Adding polarimetric decomposition features further improved CNNR's R² by 0.1524 and decreased RMSE by 0.0019 cm³/cm³ on average.

Conclusions:

• The proposed CNNR model significantly improves soil moisture retrieval accuracy compared to SVR and GRNN.
• Integrating polarimetric decomposition features from Sentinel-1 PolSAR data enhances the performance of the CNNR model.
• The combined use of Sentinel-1 and Sentinel-2 data offers a robust approach for accurate soil moisture monitoring in vegetated areas.