ADU-Net: Semantic segmentation of satellite imagery for land cover classification

Semantic Segmentation is an important problem in many vision related tasks. Land use and land cover classification involves semantic segmentation of satellite imagery and plays a vital role in many applications. In this paper, we propose an extended U-Net architecture with dense decoder connections and attention mechanism for pixel wise classification of satellite imagery named Attention Dense U-Net (ADU-Net). We further evaluate the effect of different upsampling strategies in the decoder part of the U-Net architecture. We evaluate our models on the Gaofen Image Dataset (GID) for landcover classification consisting of five classes: built-up, forest, farmland, meadow and water. The experiments on the GID dataset show better performance than the previous approaches. Our proposed architecture delivers more than 4% higher mIoU and F1-score than the baseline U-Net. Moreover, our proposed architecture achieves an F1-score of 87.21% and mIoU of 77.66% on the GID dataset. Our evaluations shows that data-dependent upsampling layer achieves higher accuracy than the Transposed Convolution, Pixel Shuffle and Bilinear upsampling layers.