IRAE-UNet: InceptionResNetV2 -Attention Encoder based UNet Semantic Segmentation of Aerial Imagery

Abstract

Remote sensing applications play a vital role in various areas such as urban planning, agriculture, and environmental monitoring. Remote sensing image segmentation, in particular, is a prominent domain that aims to address the challenges in these applications. Deep learning has significantly improved the efficiency and accuracy of remote sensing image segmentation by automating the identification of regions of interest. However, most existing methods struggle with capturing both global and local information in the images, which is crucial for accurate pixel classification. To overcome this limitation, this paper presents an enhanced version of the U-Net architecture that incorporates the InceptionResNetV2-Attention based encoder. This proposed method effectively combines the strengths of the Inception and ResNet architectures, along with the attention mechanism. The efficacy of the proposed network is verified using two publicly available datasets. The Semantic Drone Dataset consists of satellite images, while the NITRDrone dataset comprises images captured from Unmanned Aerial Vehicles (UAVs). The results demonstrate that the proposed architecture performs well on imagery obtained from different platforms, achieving a dice-coefficient of 85.04% and 88.70% for each dataset respectively, outperforming other networks.