The present invention relates to a small sample remote sensing image scene classification method based on an embedding smoothing graph neural network, which method belongs to the field of remote sensing image recognition. The method comprises: first inputting scene images into an embedding learning module, and extracting scene embedding features by means of a convolutional neural network; then introducing embedding smoothing into scene classification, and capturing the similarity and difference between the embedding features under an unsupervised condition, thereby improving the distinguishability between the embedding features, expanding a decision-making boundary, and reducing the effect of irrelevant features; in addition, constructing a graph matrix by means of an attention mechanism and by using a task-level relationship, associating a target sample with all samples in a task, and generating relationship representations having a high distinguishability between different scene categories; then, constructing graphs according to intrinsic relationships between the samples; and a label matching module iteratively generating prediction labels of samples in a test set according to the constructed graphs by means of transductive learning, until an optimal solution is obtained. By means of the present invention, accurate image classification can be realized.