Next steps for artificial intelligence in the remote sensing digital evolution
October 29, 2019
October 29, 2019
How deep learning will optimize the value of your data and tackle problems that were previously out of reach
Artificial Intelligence (AI) and Machine Learning (MI) are the fundamental building blocks of deep learning and are garnering increased attention in the geomatics community. In fact, leading research and advisory company, Gartner, has named AI as one of five emerging technology trends with transformational impact. So, it’s no surprise that AI and deep learning were both part of the buzz at the GISTAM—the annual conference of Geographical Information Systems Theory, Applications, and Management. I attended this year’s GISTAM where I met fellow practitioners and researchers from around the world who shared the latest geo-spatial ideas and innovations. Using technology, we are breaking down traditional mapping barriers.
With my focus on satellite technology—and how we use it to solve critical data problems for our clients—I was particularly interested in how deep learning technology advances were being implemented to the geomatics community.
For those who are new to the term deep learning, it is a specialized area of machine learning that draws on the science of neural networks. While that doesn’t mean we’ll have robots with human-like brains, it does mean we will be able to train algorithms to do what we do with geo-spatial imagery and elevation data faster and more accurately. As a result of GISTAM, Stantec’s Remote Sensing Center of Excellence is actively testing and evaluating several deep learning avenues to bring the best fit to our team (ENVI Deep Learning, ESRI ArcGIS Pro, Trimble eCognition).
The human brain has inspired new ways to extract information about our environment from earth observation data.
We use object-based image analysis (OBIA) for things like target vegetation species identification for monitoring ecological restoration and generating detailed land use cover maps. This allows Stantec to provide detailed maps and targeted statistics that our clients require. OBIA is great at what it does, but it also requires many user adjustments for high accuracy output. Deep learning operates more independently and can handle variations of similar target features. Everything we do with OBIA can be used to develop and create training datasets for deep learning analysis. The hope is that we can provide minimal training to deep learning software to delineate and quantify our client’s interest while reducing project costs and increasing mapping accuracy.
We just completed a very promising study in partnership with Baylor University. The study identified ephemeral stream locations that traverse pipeline rights-of-way using high resolution aerial photos, LiDAR elevation data, and deep learning. This has cost-saving potential to our clients who contract companies to walk pipeline corridors to manually identify stream crossings. This is an extremely expensive method to collect stream data. We plan to use pre-existing geo-spatial information and deep learning techniques to reduce information gathering budgets.
We are looking at deep learning for large-scale projects where target features are difficult to identify or there is a high degree of repetition to isolate surface features of interest. By incorporating deep learning into our remote sensing solutions, we’ll help our clients tackle problems that were previously out of reach. Deep learning will cover larger areas, decrease contract time, and provide clients more accurate data which enables them to make better-informed decisions for their projects.
Deep Learning is advancing and improving our current Remote Sensing services: PipeWATCH, WireWATCH, and ExtractX.