The paper presents QuadTree-Based Graph Convolutional Networks (QGCNs), which use a quadtree probability model that is first applied to small object segmentation in synthetic aperture radar (SAR) images for the case of insufficient training data. QGCNs are ensemble architectures that consist of a pre-trained encoder that forms SAR image pixel features and a neural quadtree, which is a graph convolutional network, a neural network analogue of a quadtree probability model. The graph network also contains a special branch pruning block and upscaled image features for highlighting similarities between pixels at different spatial resolutions. QGCNs with U-Net as an encoder are used to segment small objects in a few real SAR images (from the Sentinel-1 and HRSID datasets). QGCNs show higher quality in the segmentation of small objects in both multi-class and two-class problems. The difference between QGCNs and U-Net in Recall value for small objects in a multi-class problem is up to \(10.32\%\) . The increase in \(F_1\) -score for QGCNs in a two-class problem is up to \(3.59\%\) .