Near-infrared spectroscopy (NIR) technology is increasingly being applied in coal quality analysis due to its rapid and non-destructive advantages. With their compact size and ease of operation, portable NIR analyzers are particularly suited for online coal quality monitoring. However, portable devices typically exhibit lower spectral signal-to-noise ratios than traditional laboratory equipment. Furthermore, existing modeling approaches often focus on individual indicators, overlooking the interdependencies between two coal quality parameters, limiting the models' robustness and prediction accuracy. To address these limitations, this paper proposes a multi-task Unet3+ network model incorporating a channel attention mechanism, aimed at processing the low signal-to-noise ratio and weak-feature spectral data collected by portable NIR analyzers, and achieving the simultaneous prediction of ash content and calorific value in coal. First, a data preprocessing method that combines second-order differentiation with Savitzky-Golay (S-G) convolution smoothing is utilized to effectively reduce noise and enhance spectral peak features, thereby improving the data quality for modeling. Next, a Unet3+ network suitable for spectral data is designed, employing encoders, decoders, and a multi-scale feature fusion module to extract shared features relevant to ash content and calorific value. A channel attention mechanism is introduced to enhance feature representation further. Finally, the model decouples the shared features via fully connected layers to independently learn the specific characteristics of ash content and calorific value, enabling joint prediction for both tasks. Experimental validation on 670 coal samples demonstrates that, compared to typical machine learning methods and conventional deep learning models, the proposed method yields root mean square errors (RMSE) of 2.590 4 and 1.176 3 for ash content and calorific value prediction, respectively. The mean absolute errors (MAE) are 1.964 4 and 0.872 6, with correlation coefficients reaching 0.944 4 and 0.874 3, significantly outperforming the comparison models. These results provide an efficient and accurate solution for the online analysis of coal quality parameters.