Urban area has rich texture and structure information, especially from high-resolution optical images. Reducing the loss of textures and structures caused by image compression is very important for remote sensing applications. In this letter, we proposed a compression scheme for urban area based on its inherent characteristic and applications. The original image is represented by a combined dictionary. The learnt atoms are propitious to texture parts representing, and the atoms generated by Gaussian derivative functions are used to enhance the edges. Sparse coding is carried out by matching pursuit algorithm. Finally, Huffman code is applied for entropy coding. The experiments show that the proposed approach outperforms some well-known schemes in terms of texture distortion and geometric distortion.