Abstract

This paper proposes the deep convolution and correlated manifold embedded distribution alignment (DC-CMEDA) model, which is able to realize the transfer learning classification between and among various small datasets, and greatly shorten the training time. First, pre-trained Resnet50 network is used for feature transfer to extract smoke features because of the difficulty in training small dataset of forest fire smoke; second, a correlated manifold embedded distribution alignment (CMEDA) is proposed to register the smoke features in order to align the input feature distributions of the source and target domains; and finally, a trainable network model is constructed. This model is evaluated in the paper based on satellite remote sensing image and video image datasets. Compared with the deep convolutional integrated long short-term memory (DC-ILSTM) network, DC-CMEDA has increased the accuracy of video images by 1.50

Keywords

Cite this article

@article{cai-5020,
  author  = {Yaoli Wang and Xiaohui Liu and Maozhen Li and Wenxia Di and Lipo Wang},
  title   = {Deep Convolution and Correlated Manifold Embedded Distribution Alignment for Forest Fire Smoke Prediction},
  journal = {Computing and Informatics},
  volume  = {39},
  number  = {1-2},
  pages   = {318--339},
  year    = {2020},
  issn    = {1335-9150},
  doi     = {10.31577/cai_2020_1-2_318}
}
Download .bib