A lightweight Large-Scale RS image village extraction method combining deep transitive transfer learning and attention mechanism

Aiming at solving the quality and efficiency problems of village extraction in large-scale remote sensing images, this paper proposes a lightweight large-scale village extraction method that integrates deep transitive transfer learning and attention mechanism. The lightweight MobileNet v2 is used as the backbone network to solve the time-consuming problem of traditional Xception backbone network. The deep and shallow features are enhanced by introducing an attention mechanism to further improve the accuracy of village extraction. The deep transitive transfer learning strategy is used to solve the problems of wrong extraction and fragmentation of extracted villages caused by insufficient sample size in large-scale extraction, and realize the effective extraction of largescale remote sensing image villages. First, pre-train the lightweight Deeplab v3 + network with the SBD dataset to obtain the SBD pre-training weights. Then, Sentinel-2 dataset and Landsat-8 dataset were used to further train the lightweight Deeplab v3 + network successively with the SBD pre-trained weights. Then the trained proposed the lightweight Deeplab v3 + network was used to extract village from large-scale RS images. The experimental results show that the algorithm in this paper can reduce the training time. The accuracy indicators OA is 98.40 %, the Kappa reaches 0.8641, are all higher than the comparison methods. In the transferability experiment of the verification model, the OA of the proposed algorithm is above 98 %, the Kappa is above 0.83. It shows that the proposed algorithm is transferable. The proposed algorithm is applied to the Liaoning Province which village scene is complex for experiment. The result shows that it can effectively extract rural villages and has a certain generalization ability and can provide support for village monitoring in large-scale areas.