Effect of dilation rate on Nested U-Net model performance in remote sensing

High spatial resolution remote sensing images contain substantial detailed multi-scale objects. Convolutional neural networks (CNNs) are not efficient enough for detecting these objects of varying sizes. Among the multitude of CNN approaches, the Nested U-Net (UNet++) model shows great potential to capture more complex details by progressively enriching highresolution feature maps. However, there is more room for improving the Nested U-Net architecture by increasing its ability to detect multi-scale objects. The nested blocks used in this architecture rely on standard convolutional layers, which are of limited efficacy in capturing pixel information. Thus, larger receptive fields are required to extract multi-scale feature information. Although many approaches are available for increasing the receptive fields in the Nested U-Net model, these methods usually make the computational efforts very heavy. Therefore, this study uses dilated convolutions in the Nested UNet architecture to broaden the receptive field without augmenting computational demand. To this extent, the paper performs experiments with different dilation rates in the convolution blocks to understand the benefits of employing dilated convolutions in Nested U-Net architecture. Experiments using two remote sensing image sets show that the Nested U-Net model with dilated convolutions performs well for images containing both visible and multispectral wavelengths. While being able to provide performance improvement, experimental results also demonstrate that only the optimal dilation rate scheme in the proposed approach is beneficial.