Here is a list of scientific paper that make use of neuronal networks to analyze satellite images:

Some of them are using pre-trained CNN like the VGG19 and VGG 16 NN available in pytorch.

There are training dataset available for land classification training:
    http://weegee.vision.ucmerced.edu/datasets/landuse.html
    https://captain-whu.github.io/DOTA/index.html

Paper:

1. Satellite Image Scene Classification via ConvNet with Context Aggregation
   https://arxiv.org/abs/1802.00631

2. Tile2Vec: Unsupervised representation learning for spatially distributed data
   https://arxiv.org/abs/1805.02855
   https://github.com/ermongroup/tile2vec

3. Very Deep Convolutional Networks for Large-Scale Image Recognition
   https://arxiv.org/abs/1409.1556

3. Mask R-CNN from facebook research
   https://arxiv.org/abs/1703.06870
   https://github.com/facebookresearch/Detectron

4. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
   https://arxiv.org/abs/1511.06434

5. Learning Transferable Deep Models for Land-Use Classification with High-Resolution Remote Sensing Images
   https://arxiv.org/abs/1807.05713

6. Machine learning based hyperspectral image analysis: A survey
   https://arxiv.org/abs/1802.08701

   Cite: "A bias-variance analysis in [201] found that random
          forest with embedded feature selection and Markov random field (MRF)
          based post-processing are best suited for hyperspectral data."

   Cite: "Wu et al. [302] showed that a convolutional RNN [332] (a network that
          has few convolutional layers followed by RNN) is better choice for spectra
          classification than LSTM and baseline CNN."

   Cite: "Generative adversarial networks (GANs) and variational autoencoders (VAs)
          look very promising for modeling unlabeled hyperspectral data."

          "Deep learning based spatial prior could turn out to be better than
           Markov random fields based prior for land cover classification.
           GANs and VAs models for hyperspectral images should also prove to be good
           for image processing tasks, such as pansharpening, superresolution, denoising, and inpainting."

   Cite: "Another kind of data fusion would be to incorporate
          non-imaging data with imaging data. For instance, information from digital maps and geotagged user
          data could be used alongside information from remotely observed hyperspectral images for land cover
          mapping of urban areas. Due to their capabilities to model semantic relationships between different
          variables, conditional random models (CRFs) could be the best for this kind of fusion of information.
          In fact, recently a higher order CRF was used to model the dependencies between land use and land cover
          of a remotely imaged area [6]."

7. DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images
   https://arxiv.org/abs/1805.06561

8. Aerial Imagery for Roof Segmentation: A Large-Scale Dataset towards Automatic Mapping of Buildings
   https://arxiv.org/abs/1807.09532

9. Land cover mapping at very high resolution with rotation equivariant CNNs: towards small yet accurate models
   https://arxiv.org/abs/1803.06253

10. High-Resolution Semantic Labeling with Convolutional Neural Networks
   https://arxiv.org/abs/1611.01962

11. Deforestation Prediction Using Neural Networks and Satellite Imagery in a Spatial Information System
   https://arxiv.org/abs/1803.02489

12. Soft Computing Techniques for Change Detection in remotely sensed images : A Review
   https://arxiv.org/abs/1506.00768

13. DOTA: A Large-scale Dataset for Object Detection in Aerial Images
    https://arxiv.org/abs/1711.10398

14. Exploiting Deep Features for Remote Sensing Image Retrieval: A Systematic Investigation
    https://arxiv.org/abs/1707.07321

    Cite: "Even  though  some  literatures  have  made  advantages  of  deep  features  for
           RS  retrieval  task,  there  still  no comprehensive research on how to optimize
           the transferability of CNN models to RS retrieval.  With this in mind, we investigate
           almost all variables concerned to the property of CNN representations on several public
           HRRS datasets and analyze the effects of each factor."

15. Large-scale Land Cover Classification in GaoFen-2 Satellite Imagery
    https://arxiv.org/abs/1806.00901

16. Visualizing and Understanding Convolutional Networks
    https://arxiv.org/abs/1311.2901

Git repos:
    https://github.com/adnanardhian/CNN-Ensemble-Classifier-Land-Use-Classification
    https://github.com/gregbehm/UC-Merced-Pretrained-CNN
    https://github.com/chrieke/awesome-satellite-imagery-competitions


PyTorch models:
    https://pytorch.org/docs/stable/torchvision/models.html

