Enhanced root zone soil moisture monitoring using multitemporal remote sensing data and machine learning techniques.

Nouraki, Atefeh; Golabi, Mona; Albaji, Mohammad; Naseri, Abd Ali et Homayouni, Saeid ORCID: https://orcid.org/0000-0002-0214-5356 (2024). Enhanced root zone soil moisture monitoring using multitemporal remote sensing data and machine learning techniques. Remote Sensing Applications: Society and Environment , vol. 36 . p. 101354. DOI: 10.1016/j.rsase.2024.101354.

Prévisualisation

PDF - Version acceptée Télécharger (3MB) | Prévisualisation

Résumé

Accurate root zone soil moisture (RZSM) estimation using remote sensing (RS) in areas with dense vegetation is essential for real-time field monitoring and precise irrigation scheduling. Traditional methods often face challenges due to the dense crop cover and the complexity of soil and climate interactions. These challenges include the coarse spatial resolution of available soil moisture products, the influence of vegetation and surface roughness, and the difficulty of estimating RZSM from surface data. Aiming to overcome these limitations, two RZSM estimation methods were developed by combining synthetic aperture radar (SAR) data from Sentinel-1 (VV and VH polarizations) and optical and thermal RS data from Landsat-8. These data sources were used in conjunction with various machine learning (ML) models such as M5-pruned (M5P), support vector regression (SVR), extreme gradient boosting (XGBoost), and random forest regression (RFR) to improve the accuracy of soil moisture estimation. In addition to RS data, soil physical and hydraulic properties, meteorological variables, and topographical parameters were selected as inputs to the ML models for estimating the RZSM of sugarcane crops in Khuzestan, Iran. This study identified the temperature vegetation dryness index (TVDI) as a critical parameter for estimating RZSM in combination with the Sentinel-1 SAR data under high vegetation conditions. In both methods, the RFR algorithm outperformed, with similar performance, the XGBoost, SVR, and M5P algorithms in estimating soil surface moisture (R² = 0.89, RMSE = 0.04 cm³cm⁻³). However, the accuracy of the RFR algorithm decreased with increasing depth for both the optical-thermal and combined SAR and optical-thermal RS data. This decrease was more pronounced in the combined approach, particularly for the root zone, where the RMSE reached approximately 0.073 cm³cm⁻³. Accordingly, the key findings demonstrated that the optical-thermal RS data outperformed the SAR RS data for retrieving RZSM in high-vegetated areas. However, combining TVDI with SAR data is a substantial improvement that opens a new path in radar-based RZSM estimation methods under high vegetation conditions.

Type de document:	Article
Mots-clés libres:	soil moisture retrieval; landsat-8; sentinel-1 SAR; machine learning algorithms; agricultural areas
Centre:	Centre Eau Terre Environnement
Date de dépôt:	08 nov. 2024 21:29
Dernière modification:	08 nov. 2024 21:29
URI:	https://espace.inrs.ca/id/eprint/15978

Gestion Actions (Identification requise)

Modifier la notice