Estimating soil water content of cotton fields using UAV-based multi-source remote sensing data fusion

Accurate soil water content (SWC) estimation is essential for optimizing irrigation strategies in cotton cultivation, especially during the flowering, boll setting, and boll opening stages, when SWC variations critically impact yield and fiber quality. Multi-source data fusion provides a powerful method for estimating SWC by effectively leveraging the advantages of different datasets to enhance estimation accuracy and robustness. However, it often fails to comprehensively account for the interactions within the soil-plant-atmosphere continuum, which limits the accuracy of the estimates. This study multidimensionally fused thermal infrared, multispectral, and meteorological data to systematically evaluate machine learning algorithms for improving SWC estimation accuracy. Joint water and nitrogen treatments, including three irrigation regimes and three nitrogen application rates, were conducted in the cotton field. SWC data were collected from six soil depths during critical cotton growth stages, and Unmanned Aerial Vehicle (UAV)-based images were also acquired. A multidimensional remote sensing feature set was constructed, containing the crop water stress index (CWSI), normalized difference vegetation index (NDVI), temperature vegetation dryness index (TVDI), and three-dimensional drought index (TDDI). Four machine learning algorithms, i.e., support vector regression (SVR), random forest regression (RFR), gradient boosting decision tree (GBDT), and categorical boosting (CatBoost), were evaluated for their performance in estimating SWC. Results showed that TVDI _{G, R} exhibited the strongest correlation with SWC (r = -0.47 ± 0.03), demonstrating heightened sensitivity at soil depths of 0 – 10 cm (r = -0.5). In addition, the CatBoost model showed superior estimation performance (R² = 0.762 ± 0.026), significantly outperforming other models. Its robustness strengthened, following proceeding cotton growth and augmenting irrigation amount, despite of demonstrating robust generalizability. In conclusion, the accuracy of SWC estimation was enhanced by integrating multidimensional indices with machine learning techniques, and a practical technical framework to support precision irrigation in arid cotton agroecosystems was developed.