CSDMS3.0 - Bridging Boundaries

Landcover and landform classification using deep neural networks

Daniel Buscombe

Northern Arizona University, United States

This clinic will introduce deep learning methods for semantic segmentation of fluvial sedimentary landforms and riparian environments, using high-resolution aerial imagery. Deep neural networks are the current state-of-the-art for discrete classification of remotely sensed imagery from Earth observation platforms. The clinic will guide users through the process of preparing training datasets, training models, and evaluation. A number of different deep convolutional neural network architectures for image feature extraction and pixel-scale classifications will be explored and compared. The clinic will use the keras and tensorflow libraries within the python programming language. This hands-on class will be taught using Google colab through a browser, with the materials hosted on github. Participants will require a working knowledge of python. Some working knowledge of machine learning would be helpful, but we will assume no prior experience with machine/deep learning, neural networks, tensorflow, or keras.

Please acknowledge the original contributors when you are using this material. If there are any copyright issues, please let us know ( and we will respond as soon as possible.

Of interest for:
  • Terrestrial Working Group
  • Coastal Working Group
  • Cyberinformatics and Numerics Working Group
  • Hydrology Focus Research Group
  • Ecosystem Dynamics Focus Research Group
  • Coastal Vulnerability Initiative
  • Artificial Intelligence & Machine Learning Initiative