In the field of remote sensing (RS) archaeology, the identification of archaeological traces is an important foundation for further archaeological analysis and understanding of the ancient landscape. However, conventional methods rely heavily on visual interpretation by archaeologists and RS specialists, which are difficult to handle the large amount of archaeological traces in massive RS data. Especially for large-scale linear archaeological sites, such as the Great Wall, it is challenging to obtain systematic and complete identification results through human–computer interpretation. Based on the theory of deep learning, we propose an improved DeepLabv3+ model and conduct a typical application of the Great Wall of the Han dynasty (GWH) in northwest (NW) China using very high-resolution (VHR) aerial imagery. The main improvements of the DeepLabv3+ model were (1) replacing the encoder module with pre-trained ResNet101 to obtain deeper features, and (2) adding the Dice coefficient to the loss function to improve the accuracy under the unbalanced condition of positive and negative samples. The experimental evaluation reveals that the improved DeepLabv3+ model achieved accuracy of 98.37%, recall of 98.92%, F1 of 98.64%, and intersection over union (IoU) of 97.32%. Compared with the original model, accuracy, recall, F1, and IoU have improved by 9.03%, 9.13%, 9.08%, and 16.22%, respectively. Furthermore, the identification results in the test area also indicate that the improved model has good transferability. This study provides an automatic approach for surveying and mapping the GWH and indicates the great potential of deep learning in the identification of linear archaeological traces in RS archaeology.