Tutorial demonstrating how to create a semantic segmentation (pixel-level classification) model to predict land cover from aerial imagery. This model can be used to identify newly developed or flooded land. Uses ground-truth labels and processed NAIP imagery provided by the Chesapeake Conservancy.
Switch branches/tags
Nothing to show
Clone or download
Latest commit 054bf54 Aug 29, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
batchai Update evaluate.md Aug 29, 2018
geoaidsvm Update 04_Apply_trained_model_in_ArcGIS_Pro.ipynb Aug 13, 2018
outputs Add files via upload Aug 13, 2018
sample_models Add files via upload Jan 18, 2018
scripts switched the order h and w Feb 27, 2018
.gitignore Initial commit Jan 18, 2018
LICENSE Initial commit Jan 18, 2018
README.md Added link to blog post Apr 8, 2018


Pixel-level land cover classification

This repository contains a tutorial illustrating how to create a deep neural network model that accepts an aerial image as input and returns a land cover label (forested, water, etc.) for every pixel in the image. Microsoft's Cognitive Toolkit (CNTK) is used to train and evaluate the model on an Azure Geo AI Data Science Virtual Machine or an Azure Batch AI GPU cluster. The method shown here was developed in collaboration between the Chesapeake Conservancy, ESRI, and Microsoft Research as part of Microsoft's AI for Earth initiative.

We recommend budgeting two hours for a full walkthrough of this tutorial. The code, shell commands, trained models, and sample images provided here may prove helpful even if you prefer not to complete the walkthrough: we have provided explanations and direct links to these materials where possible.

How to Get Started

The training and evaluation steps of this tutorial can be performed on either:

  • an Azure Geo AI Data Science VM
    • Train a model on a data sample using Jupyter notebooks
    • Deploy the trained model directly in ESRI's ArcGIS Pro
  • an Azure Batch AI GPU cluster
    • Set up your cluster and submit jobs to it from your command line
    • Learn how to scale to large clusters for faster training on larger datasets
    • (Optional) After training, you may download your model and deploy it in ArcGIS Pro on a Geo AI DSVM (see provisioning instructions to get started)
Geo AI DSVM Batch AI
Create a Geo AI Data Science VM Create a Batch AI cluster
Train a model in a Jupyter notebook Train a model on the Batch AI cluster
Evaluate the model using a Jupyter notebook Evaluate the model using a GPU cluster
Deploy your model in ArcGIS Pro Learn how to scale up training

Sample Output

This tutorial will train a pixel-level land cover classifier for a single epoch: your model will produce results similar to bottom-left. By expanding the training dataset and increasing the number of training epochs, we achieved results like the example at bottom right. The trained model is accurate enough to detect some features, like the small pond at top-center, that were not correctly annotated in the ground-truth labels.

Related materials


This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.