Digital surface model extraction with high details using single high resolution satellite image and SRTM global DEM based on deep learning

The digital surface model (DSM) is an important product in the field of photogrammetry and remote sensing and has variety of applications in this field. Existed techniques require more than one image for DSM extraction and in this paper it is tried to investigate and analyze the probability of DSM extraction from a single satellite image. In this regard, an algorithm based on deep convolutional neural networks (CNN) is designed. In the proposed subject, firstly, some preprocessing such as dividing the satellite image into smaller images, localizing the height values and data augmentation are applied in order to prepare data to enter the network. The proposed CNN network has an encoder-decoder structure in which, different and effective features in different scales are extracted in the encoder stage and the generated features are fused to estimate height values by presenting an effective procedure in the decoding stage. Subsequently, the ground and non-ground pixels are separated and height values of the non-ground objects are extracted. The final DSM is obtained by adding the non-ground pixels with height information to the SRTM digital elevation model (DEM) with 30 meter pixel size. The proposed algorithm is evaluated using the satellite images and their corresponding DSMs. Analyzing the estimated small height images using the proposed CNN indicated 0.921, 0.221 and 2.956m on average for relative mean error (ER), logarithm mean error (EL) and root mean squared error (ERMSE), respectively. Moreover, analyzing the final seamless DSMs indicated 4.625 on average for ERMSE.