It is crucial but challenging to detect intermediate or end products promptly. Traditional chemical detection methods are time-consuming and cannot detect mineral phase content. Thermal infrared hyperspectral (TIH) technology is an effective means of real-time imaging and can precisely capture the emissivity characteristics of objects. This study introduces TIH to estimate the content of potassium salts, with a model based on Competitive Adaptive Reweighted Sampling (CARS) and Partial Least Squares Regression (PLSR). The model takes the emissivity spectrum of potassium salt into account and accurately predicts the content of Mixing Potassium (MP), a mineral mixture produced in Lop Nur, Xinjiang. The main mineral content in MP was measured by Mineral Liberation Analyzer (MLA), mainly including picromerite, potassium chloride, magnesium sulfate, and less sodium chloride. 129 configured MP samples were divided into calibration (97 samples) and prediction (32 samples) sets. The CARS-PLSR method achieved good prediction results for MP mineral content (picromerite: correlation coefficient of correction set (Rp2) = 0.943, predicted root mean square error (RMSEP) = 2.72%, relative predictive deviation (RPD) = 4.24; potassium chloride: Rp2 = 0.948, RMSEP = 2.86%, RPD = 4.42). Experimental results convey that TIH technology can effectively identify the emissivity characteristics of MP minerals, facilitating quantitative detection of MP mineral content.