Building extraction from VHR remote sensing imagery by combining an improved deep convolutional encoder-decoder architecture and historical land use vector map