Community
Select to view content in your preferred language

How to get started with GeoAI in ArcGIS - A practical guide for students

410
0
Wednesday
CanserinaKurnia
by Esri Regular Contributor
Esri Regular Contributor
2 0 410
Wednesday

Many students are interested in trying GeoAI capabilities in ArcGIS.  Some have only a few hours to spend, others plan to use it for class projects, and some aim to use it for a longer research project.

The goal of this blog is to guide you to take your first steps into the world of GeoAI using ArcGIS, focusing specifically on deep learning.

 

What and Why GeoAI

GeoAI—short for Geospatial Artificial Intelligence—is revolutionizing how we analyze spatial data. By combining GIS with machine learning, deep learning, and other AI techniques, GeoAI empowers GIS to extract features, uncover patterns, make predictions, and automate workflows, and boost productivity in ways never before possible.

Key benefits examples:

  • Automation of spatial tasks like feature extraction from imagery – such as extracting building footprints from imagery
  • Performing Analysis like detecting vegetation encroachment to the electrical utility from point clouds or identifying wildfires risk zones
  • Increasing productivity automates repetitive tasks, reducing human errors and more consistent updates and results – such as mapping land use changes

GeoAI technology integrates well with GIS.   Most of the time the result from GeoAI becomes an input in GIS analysis.  A good example is in this tutorial:

 

In ArcGIS, GeoAI can be applied to:

  • Imagery analysis: e.g land cover classification, object detection, change detection,
  • Vector data modeling: e.g clustering, regression, classification, time series forecasting
  • Text and tabular data mining: e.g entity extraction, address correction,
  • 3D and time-series forecasting: 3D feature extraction

 

In ArcGIS, GeoAI is capability that can you access across the platform: ArcGIS Pro, ArcGIS Online, ArcGIS Enterprise, and through the ArcGIS Python API (arcgis.learn) and deep learning studio.

IMPORTANT: To use GeoAI capabilities in ArcGIS Pro, an additional installation of deep learning libraries is required. The deep learning libraries installers for ArcGIS can be downloaded from GitHub. Ensure that the installer version matches your ArcGIS Pro version. Refer to this tutorial for step-by-step installation instructions.

 

Getting started pathway

There are several options to get started performing GeoAI in ArcGIS.   Take a look this diagram:

Options to perform GeoAI in ArcGISOptions to perform GeoAI in ArcGIS

 

Use Pretrained Models

Using pretrained GeoAI models is the simplest way to begin, by using the models that have been trained.  With ArcGIS pretrained models, you do not need to invest time and effort into training a deep learning model—tasks that are both time-consuming and computer resource-intensive.

There are more than 100 pretrained models available in the ArcGIS Living Atlas, each designed for various applications.  The collections are ready-to-use, in the form of deep learning packages (dlpk) that you can download or use directly in ArcGIS.   The collection continues to grow as new models are added regularly. 

 

(view in My Videos)

 

Accessing deep learning package from ArcGIS Living Atlas of the World

In general, there are two types of pretrained models:

 

Task-Specific Models

These models are built for a single, well-defined task like detecting buildings or land cover classification.  These models are purpose-built for geospatial analysis and closely aligned with real-world mapping needs.  Characteristics of task-specific models:

  • Higher accuracy
  • Support wide variety of geospatial formats
  • All AI tasks: object detection, pixel classification, change detection, and more
  • Smaller models, more efficient and can run with lower computing requirements
  • Can be fine-tuned and retrained to improve the result

Examples:

Tutorials:

 

Generalized Models

These models can handle many tasks often through flexible natural language prompts such as detecting red cars or airplanes.  These models are designed for board, non-spatial image understanding tasks.  Characteristics of generalized models:

  • May compromise on precision
  • Typically work only with true color pictures, RGB imagery
  • Object classification and segmentation
  • Larger model, computing-intensive, requiring more powerful GPUs or cloud services
  • Mostly static and cannot be fine-tuned or improved

Examples:

Tutorials:

Note: You can use other deep learning models to improve the results. 

In addition to task-specific and generalized models, the pretrained models in Living Atlas also include several deep learning packages that serve as a bridge to the AI community's resources on the Hugging Face Hub (HF).

Examples:

View the complete list of pretrained models in ArcGIS Living Atlas

 

Tips and Best Practices

Each pretrained model provides metadata. It is important to review the information provided in the metadata before using a pretrained model:

  • Model description
  • Guides for using the model
  • Details on the data and resolution used to train the model
  • Information regarding fine-tuning capabilities
  • Geographic relevance of the model's performance
  • Model architecture, accuracy metrics, and sample results

 

Fine-tune the results by retraining the model using Transfer Learning

It is quite rare that you get satisfactory results just by running or using the pretrained model once.  You may reach a certain level of good results, but you may want to enhance or improve the results.   One method is to retrain the model using transfer learning method   This mostly applies to task-specific models.  Always check the metadata of the model, if that model can be retrained.  You can use ArcGIS Pro tools to retrain the model. This is the workflow to retrain the model in general

  • Collect more samples in your study area and label them appropriately
  • Use the tool Export to Deep Learning to convert the samples to chips
  • Use tool to train deep learning to include both the pretrain model and the additional sample chips
  • Test the model by inferencing.

The process of retraining the model is following the same workflow as you train a new model, however it only takes a small amount of samples. 

Check this tutorial to learn how to retrain a model using transfer learning method:

Improve deep learning model with transfer learning

 

Train your own model

If none of the pretrained models work for your case, you can train your own model.  ArcGIS Pro provides end-to-end tools to train your own model, from collecting samples, export them to deep learning, training the model by choosing variety of deep learning frameworks models and finally inferencing to test re model.  These are the same tools that you use when you fine-tune an existing trained model.

Tutorials:

 

Applying GeoAI with Python libraries

For those who are familiar with python programming, you can apply GeoAI though python coding.  You can arcgis.learn library in ArcGIS and combining it with other libraries from open science.  This is actually give the most flexibility to apply GeoAI for many different modalities. 

Take a look of a few sample notebooks for performing GeoAI in ArcGIS:

 

Conclusion

GeoAI is an emerging technology that encompasses a wide range of workflows, methods, data, and experiences. Beyond deep learning, it includes various machine learning models and algorithms for tasks like classification, prediction, clustering, and time series forecasting.

The most important thing for the students is to understand the fundamentals of how to perform GeoAI and on your creativity to keep trying to get the results that you need. Understanding and being critical in reviewing the results are equally important.

Check these resources to familiarize yourself with GeoAI examples and tutorials.

More:

 

Contributors
About the Author
Canserina Kurnia is a GIS professional with over 25 years of experience. She currently holds the position as a Senior Solution Engineer at Esri, at their headquarter office in Redlands, California. Her main role is to provide technical advices and assistance to universities globally, in advancing their GIS technology for teaching and research.