Abstract: This paper, based on deep learning methods, adopts the open-source TensorFlow framework and employs the DeepLabv3+ network model with excellent segmentation accuracy to complete the automatic identification and extraction of building contours from remote sensing images.The automatically extracted vector building patches boast complete boundaries and clear details, which can objectively and accurately reflect the spatial distribution and agglomeration characteristics of various types of buildings in the study area.Compared with the traditional manual vectorization approach via visual interpretation, this scheme can shorten the data production cycle and improve acquisition efficiency to a certain extent, providing a feasible automated way to obtain basic data of seismic disaster-bearing bodies.The simultaneously extracted building footprint data boasts high timeliness and fine spatial granularity.It accurately captures architectural details of residential settlements, effectively compensating for the inherent drawbacks of traditional base geographic data, such as insufficient accuracy and lagged update cycles.This dataset provides accurate, scientifically rigorous foundational data support for earthquake disaster loss pre-assessment and rapid post-earthquake disaster evaluation.