Active learning based semantic segmentation for extraction of minute objects from multispectral satellite images

In this manuscript, an attention based Unet (A-Unet) has been proposed for semantic segmentation of enormous remotely sensed images for the extraction of minute objects like electrical substations. For this purpose, a pre-trained deep convolutional neural network (DCNN) has been used in every encoder-decoder level over the A-UNET. It is infused with an intelligent active learning approach for retraining and thereby leveraging the segmentation model for detecting smaller objects from large-scale images. The experimentation has been carried out in three phases: pre-processing the images, training of an A-Unet framework, and active learning based retraining methodology to achieve improvement over the overall learning curve. The results validated over two state-of-the-art remote sensing datasets (of electrical substation and ships) show efficient performance in segmenting the necessary small object boundaries. A mean intersection over union (mioU) calculated for this purpose show promising results in favour of the proposed approach in comparison to the state-of-the-art methodologies.