To address the online monitoring of molten salt temperature in the rare earth electrolysis process, this paper proposes a rare earth molten salt temperature detection model (MSTDM) based on multi-color space features and data regression neural networks. The model initially preprocesses the collected molten salt temperature images using techniques such as the Gaussian Mixture Model (GMM) and median filtering, followed by color space transformation. Subsequently, 19 hue features used for temperature quantification are extracted, including the channel responses of RGB, HSV, Lab, and Gray color spaces. Finally, a novel data regression neural network (CNN-BiLSTM-Attention) is employed to learn the long-term dependencies between molten salt hue features and temperature, predicting future molten salt temperatures. Experimental results indicate that using the GMM algorithm to filter out flame interference, introducing novel temperature prediction features, and constructing the CNN-BiLSTM-Attention network all enhance the performance of MSTDM in molten salt temperature prediction tasks. The model achieves a mean absolute error of 3.16 °C in predicting molten salt temperatures at rare earth smelting sites. This research not only provides an effective solution for online monitoring of molten salt temperatures in the rare earth electrolysis process but also offers a reliable reference for temperature monitoring in other fields.