Please enjoy the Esri Pipeline recording of the webinar titled, "Merging Location and Artificial Intelligence," held on April 7, 2020.
Below you will find select questions and answers from the webinar session. Thank you.
QUESTION: Is the AI training/learning one task at a time or can you nest or build the learning tasks?
ANSWER: It depends on which learning tasks you would like to be coupled - training a model to classify various animals like cats and dogs can be single model (you don’t need a cat model, a dog model, etc) - but for a model that does two different kinds of tasks, generally the answer is no. Model architecture selection is important for a given use case and the architectures can’t be fused, so generally tasks of a different kind should be separate.
For example, pixel classification at high accuracy has been achieved efficiently using UNet and object detection has great performance with ResNet. It’s not possible to fuse these architectures and have a single model do the tasks, but you could sequentially call the trained models to perform their individual tasks such as using an object detection model to detect a well pad and then point a pixel classification model at the subset of the image that is the well pad to identify asset equipment like tanks and batteries.
QUESTION: Is there a pool of GeoAI training script that can be used, or will users need to program all unique scripts all the time?
ANSWER: It depends on what is meant by “training script.” There are three things to note:
- The process of actually training a model is two lines of code in the ArcGIS API for Python – choose a model architecture, call the model.fit method on the instantiated architecture and that will output a trained model.
- The training process in ArcGIS Pro requires no scripting – Export Training Data for Deep Learning, Train Deep Learning Model, and Detect Objects Using Deep Learning/Classify Pixels Using Deep Learning are all ArcGIS Pro geoprocessing tools and these could be put into a workflow that is automated if your data supports that ease of use.
- There are samples for the ArcGIS API for Python at developers.arcgis.com, including a GitHub repository full of Jupyter Notebooks. Not all of those are machine learning, but several of them are. With that said, adapting these to your dataset and parameters, your model architecture, your use case will generally always be necessary.
QUESTION: Are there free online tutorials for this workflow?
ANSWER: Yes! Check out the ArcGIS Spatial Data Science MOOC on the Esri Academy. This has a lot of great resources for the AI/ML process from data engineering to model development to data visualization and sharing. You can also find samples for the Python API at developers.arcgis.com, several of which demonstrate our machine learning capabilities. For the specific use cases demonstrated in the video, there are no tutorial-style walkthroughs, but there are articles that describe the workflows in technical detail at medium.com/geoai.
QUESTION: Are other pipeline data models like PODS supported?
ANSWER: Yes, any of the industry data models like PODS and UPDM are supported so long as they are spatialized
QUESTION: Is there a way to get an evaluation license for "ArcGIS Image Analyst" extension?
ANSWER: Yes, we can offer a standard evaluation license for Image Analyst; please reach out Jeff Allen for more details.
QUESTION: What is the role of Linked data in spatial analysis?
ANSWER: Linked data is a technical term that refers to a specific method of publishing interlinked datasets on the web to enable more effective semantic querying, allowing applications to more effectively utilize data pieces by connecting them to other pieces of data and thus infer meaning or logic from those pieces of data. A good example of this (and a common implementation of linked data) would be any knowledge graph.
In spatial analysis, you might use a graph data structure that represents a linked dataset to connect semantic meaning with location. This could be as simple as associating XY coordinates with a place-name or as complex as using a combination of location, communication metadata, and financial information to link establish relationships between people (as is a common use case in the defense/intelligence/law enforcement community). This is sometimes called link analysis or network analysis and spatial components underly that analytical process.
QUESTION: Would it be possible to automate HCA analysis using machine learning?
ANSWER: Yes, machine learning can be used to detect buildings, and other locations, where pipeline releases could have greater consequences to health and safety or the environment.
QUESTION: Are these models available for us to use?
ANSWER: The model architectures have been constructed and are available for use in the ArcGIS API for Python in the arcgis.learn submodule for you to train on your own datasets. However, the models that have been trained to perform specific tasks like detecting well pads or constructing road networks are not freely available. If you want Esri to provide a trained model to perform a specific task, contact us and we can discuss your requirements - we commonly provide this service to our clients.
QUESTION: Can you please share this presentation with me?
QUESTION: What's the better UPDM or PODS?
ANSWER: Esri pipeline data management tools are pipeline data model agnostic. There are many reasons why an operator may choose one industry standard model over another. Please reach out to Jeff Allen for more information on how these models are being successfully deployed within the industry.
QUESTION: Does Esri have a facility data model; e.g., managing building, asset, retirement, etc., where I can download the xml?
ANSWER: Yes, the Esri facility data model is UPDM and it can be downloaded here
QUESTION: Do you have any examples of successfully applying these techniques to seabed data? The results of the multibeam surveys produce extensive point sets. We deal only with submarine pipelines but still have large datasets.
ANSWER: Although we don’t currently have an example of applying these machine learning algorithms to subsea multibeam survey data we have had success in applying these machine learning techniques to a wide array of lidar and other sensor-based datasets. Please reach out to use if you would like to explore this use case further.