Aerial Image Segmentation Using Deep Learning

Abstract

Abstract Image segmentation of aerial images using deep learning has gained significant attention due to its potential for extracting valuable information from high-resolution imagery. This study focuses on the application of deep learning techniques for image segmentation in the context of aerial images. Specifically, popular architectures such as Unet, PSPNet, and LinkNet are utilized with different Feature Extraction Networks including EfficientNet-B4 and ResNet50. The models are trained and evaluated on aerial imagery of Dubai obtained by MBRSC satellites dataset, and the results are assessed using Intersection over Union (IoU) metrics for training and validation sets. The findings reveal that Unet with EfficientNet-B4 achieves the highest IoU scores, with a training IoU of 0.65708 and a validation IoU of 0.64002. PSPNet and LinkNet also demonstrate competitive performance, with EfficientNet-B4 as the preferred Feature Extraction network. These results highlight the effectiveness of deep learning approaches for aerial image segmentation and provide valuable insights for selecting suitable models and architectures for this task.