Hyperspectral imaging is an image obtained by combining spectral detection technology and imaging technology, which can collect electromagnetic spectra in the wavelength range of visible light to near-infrared. It is an important research content in the field of ground observation in hyperspectral remote sensing. However, hyperspectral image face significant challenges in classification task due to their high spectral dimensions, lack of labeled samples, and strong correlation between bands. In order to fully extract features from both spectral and spatial dimensions and improve classification accuracy in the case of limited training samples, a multiscale dilated attention network is proposed for hyperspectral image classification. First, a three-dimensional convolutional layer is used to extract the shallow features of the image. Then, a multiscale dilated attention module is proposed by combining dilated convolution and channel attention. Using ordinary convolution and dilated convolution to form different receptive fields. Channel attention is used to remodel the obtained multiscale features, enhancing the inter-channel correlation. After that, a multiscale spatial-spectral attention module is constructed using multiple asymmetric convolutions to obtain spatial and spectral attention features at different positions, further enhancing important feature suppression over non-important features. Finally, using softmax to classify the obtained features. Using Indian Pines, Pavia University, KSC and University of Houston as experimental datasets, the overall classification accuracy of this paper’s method achieved 98.97%, 99.14%, 99.45%, and 98.56% respectively, using only 5%, 1%, 10%, and 10% of training samples per class. Compared with seven advanced classification methods, the experimental results show that the proposed method can achieve the highest classification accuracy with limited training samples.