Neural quadtree and its applications for SAR imagery segmentation

The paper considers the neural ensemble neural network architecture that uses a quadtree model for SAR (Synthetic Aperture Radar) image segmentation under the lack of training data. The Neural Quadtree network (NQN) consists of segmentation network forming the image pixels features and a graph convolution network with the special branch pruning block establishing the spatial and hierarchical connections between pixels. The NQN is used for segmenting of several SAR images that differ a lot both in presented surfaces and characteristics (Sentinel-1, ESAR (Experimental SAR), HRSID (High-Resolution SAR Images Dataset)). A comparison was made of the results of processing images of NQN and a conventional quad-tree using a common U-Net network segmentor. The NQN demonstrates the higher quality in target detection in comparison with a conventional quadtree. The difference in Recall values for such objects classes between NQN and quadtree ranges from 2.13% to 11.63%.