Comparison Between UNet, Modified UNet and Dense-Attention Network (DAN) for Building Extraction from TripleSat Imagery

摘要

Building extraction from high resolution remote sensing imagery is of great importance for land use analysis, urban planning, and many other applications. Notably convolutional neural networks (CNN) have shown significant advantage over traditional methods for this task. Among various CNN models, UNet has gained popularity due to its simplicity, efficiency and robustness while many modified versions have been proposed. More recently, a model called the dense attention network (DAN) based on DenseNets and attention mechanism was proposed. This model achieved good performance in building extraction from very high resolution imagery. Based on these developments, in this paper, we compared three architectures (UNet, modified UNet (with residual blocks and recurrent feature), and DAN) for building extraction in Kuala Lumpur, Malaysia using 0.8m TripleSat imagery. For the modified UNet. skip connections were implemented in each encoder blocks to mix features of different levels. Output was multiplied to input and feed to the same UNet again. The comparison results showed that the modified UNet achieved the highest Fl-score. while the DAN achieved average higher Fl-score than the UNet. But DAN had the highest accuracy for validation patches with large buildings.