Accurately acquiring soil moisture content (SMC) and its spatial distribution is of great significance for water-saving irrigation and sustainable agricultural development in arid regions. However, the complex soil background noise and weak moisture absorption features in the Xinjiang region severely restrict the accuracy and reliability of remote sensing inversion for SMC. To address the challenges of difficult feature capture and low estimation accuracy in soil moisture monitoring, this study utilized GF-5 satellite hyperspectral data and ground-measured SMC data. First, the effects of Fractional-Order Differentiation (FOD) at orders 0–2 (with a 0.2 step size) on spectral moisture response were systematically evaluated. Next, 60 dual-band spectral indices (DBIs) were constructed from full-band combinations under the optimal differentiation orders, and highly correlated indices were selected as candidate features. Finally, three variable screening methods were coupled with three machine learning models to construct nine SMC inversion schemes, and the optimal model combination was employed to map the spatial distribution of SMC in the study area. Results showed that FOD at orders 0.8–1.2 effectively enhanced spectral responses at soil moisture absorption bands, the introduction of DBI concentrated high-correlation band combinations in moisture-sensitive regions, and the optimal scheme (BSS-PLSR) demonstrated good predictive performance and stability. These findings provide data support for precision irrigation decision-making and soil moisture management in arid farmlands.