Building Extraction in Multitemporal High-Resolution Remote Sensing Imagery Using a Multifeature LSTM Network

Abstract

Inspired by recent promising developments by deep learning networks, this letter presents a novel approach based on a multifeature long short-term memory (LSTM) network with an optimal unit number for extracting buildings from multitemporal high-resolution optical satellite imagery. The algorithm first extracts specified multifeature data of buildings, including spectral, shape, texture, and indices from multitemporal high-resolution imagery. Next, a designed LSTM network architecture with an optimal unit number is trained using the multifeature data to extract buildings at the pixel level. The proposed network is compared with popular deep learning-based networks, including VGG, U-Net, and ResNet. Finally, postprocessing (e.g., morphological algorithm) is employed to optimize the classification results produced by deep learning. The approach is validated on a newly created data set using Gaofen-2 (GF-2) imagery that contains various buildings on a campus with different materials, shapes, sizes, heights, and directions. Experimental results indicate that the proposed approach outperforms current deep learning-based methods for building extraction and has a potential for practical applications.