Clouds exert a significant influence on infrared radiation, making cloud detection a crucial step in the application of infrared hyperspectral data. In particular, water vapor interference and the limited accuracy in high-cloud identification constitute two key challenges for ground-based infrared hyperspectral cloud detection. Traditional threshold-based cloud detection methods are difficult to adapt to different locations and dynamically changing atmospheric conditions,while machine learning methods can achieve cloud detection with higher accuracy, greater robustness, and improved automation. Building on the advantages of machine learning, observational data from the atmospheric sounder spectrometer by infrared spectral technology (ASSIST), collected at Lijiang (Yunnan), Motuo (Xizang Autonomous Region), and Ritu (Xizang Autonomous Region) in China, are used to analyze the spectral differences between sunny and cloudy conditions in this study. Based on these differences, a spectral feature enhancement-driven machine learning method for cloud detection is proposed. Finally, by incorporating synchronous observations from lidar, meteorological stations, and all-sky imagers, the proposed method is systematically evaluated under different relative humidity (RH) and cloud base height (CBH) conditions. The experimental results show that the consistency between the results obtained by the proposed method and lidar-based detection is as high as 97.61%. Under different RH conditions, the proposed method outperforms the method based on original spectral features. Notably, when ${\text{RH}} > 70{\text{%}} $, the accuracy of clear-sky spectral identification improves significantly: increasing from 86.01% to 91.89%. Similarly, under different CBH conditions, the proposed method also exhibits superior performance compared with the method in which original spectral features are used. In particular, the accuracy improvements are especially notable when identifying mid-level clouds with ${\text{3 km}} < {\text{CBH}} \leqslant 5{\text{ km}}$, as well as high-level clouds with ${\text{CBH}} > 5{\text{ km}}$. When ${\text{3 km}} < {\text{CBH}} \leqslant 5{\text{ km}}$, the accuracy increases from 95.45% to 98.64% and when ${\text{CBH}} > 5{\text{ km}}$, the accuracy improves from 87.5% to 91.67%. The proposed method significantly enhances the automation and accuracy of cloud detection, thereby providing higher-quality fundamental datasets for supporting subsequent applications such as radiative transfer simulation, remote sensing parameter retrieval, and data assimilation in numerical weather prediction (NWP) models.