Abstract: High-precision online detection of ash content in flotation tailings is essential for intelligent coal preparation, but conventional vision-based methods are susceptible to interference from water films and metamerism. To address this issue, the diffuse-reflectance optical response mechanism of flotation-tailings sediment was analyzed, and the feasibility of predicting ash content from sediment hyperspectral images was demonstrated. A dataset spanning the full ash-content range was constructed by artificially blending raw samples from an operating coal preparation plant at graded proportions while preserving their original mineral sliming characteristics. A flotation-tailings ash content prediction method integrating spatial semantic segmentation with a deep temporal network was proposed. A full-band U-Net segmentation model and a morphological truncation algorithm were developed to extract sediment regions. Isolation Forest, multiplicative scatter correction, and continuum removal were used for preprocessing, and a dual-channel feature tensor was constructed. A concentration-prior-integrated 1D-RSL ash content prediction model was then constructed by combining a One-Dimensional Residual Network (1D-ResNet), Squeeze-and-Excitation (SE) channel recalibration, and Long Short-Term Memory (LSTM) to extract local interband features and long-range dependencies and achieve sample-level ash content prediction. Test results showed that the coefficient of determination and root mean square error of the model predictions were 0.952 and 1.95%, respectively, and its prediction error was lower than those of models such as partial least squares regression and support vector regression. A complete detection process took approximately 48.2 s and met the requirements for dynamic monitoring of flotation-tailings ash content in industrial production.