To solve the problems of strong subjectivity, long time-consumption and sample destructiveness of the existing umami intensity detection methods, deep learning and machine learning algorithms combined with hyperspectral imaging technology were used to establish a fast and nondestructive detection method for grass carp umami intensity. After collecting the hyperspectral data of grass carp, the spectral feature wavelengths were selected using competitive adaptive reweighted sampling method, and the Gaussian-weighted multi-head attention network (GMANet) was developed and support vector machine regression (SVR), partial least squares regression (PLSR) and other machine learning algorithms were used to establish and optimize the grass carp umami detection model. The results showed that the root mean square error of prediction and the coefficient of determination of prediction of GMANet network were 0.0082 and 0.8844, respectively, which were better than the optimal modeling method SVR in traditional machine learning, whose root mean square error of prediction and the coefficient of determination of prediction were 0.0077 and 0.8188, respectively. The study shows that hyperspectral technology has a large application prospect in the direction of umami intensity detection, and the GMANet network can make full use of the spatial image and spectral information of the samples, which provides a new method for the subsequent application of hyperspectral image detection.