<p>The Qingshui River Basin is the largest tributary entering the Huanghe (Yellow) River within Ningxia of China, contributing about 49% of the sediment load from this section. In this basin, the suspended sediment concentration (SSC) exhibits a short-term variability with values rising by orders of magnitude within hours. This phenomenon is primarily driven by the precipitation-induced runoff, which induces the surface soil erosion, sediment transport and the local bed erosion. Such rapid fluctuations of SSC are difficult to be captured using manual sampling or satellite remote sensing methods. To address this issue, a classification-based inversion model was developed in this paper using the ground-based in-situ hyperspectral data from the Wangtuan station beside Qingshui River between March and September 2024, covering both normal and flood periods. At the beginning, we utilized the K-means clustering algorithm to classify these water spectral characteristics into three types. These water types exhibited different spectral characteristics in the position and number of their reflectance peaks. Subsequently we built empirical and machine learning models for each water type to retrieve SSC. By comparison to Global models, the classification-based inversion models substantially improved inversion performance: a linear empirical model performed the best for low SSC water type (<i>R</i>²=0.92), while support vector regression and random forest achieved an <i>R</i>² of 0.92 under medium and high SSC conditions, respectively. Therefore, this study we conducted in this paper achieved a high inversion accuracy of SSC especially for medium and high SSC water types. Our results revealed that the ‘classify first, then model’ method is reliable for estimating SSC of high-sediment water bodies in the Qingshui River Basin, and enhances the feasibility of ground-based in-situ hyperspectral monitoring technology for refined Huanghe River watershed management.</p>