Pre-trained deep learning models are now available from Esri. These models make the benefits of deep learning more accessible to a wider audience of GIS users and analysts who may not have the technical know-how, training data, or computing resources to build their own models from scratch.

Deep Learning

Deep Learning is a type of artificial intelligence that relies on developing and training a neural network model with many layers. Most often this involves using massive amounts of labeled training data. The training process also frequently requires huge quantities of computing resources. Once a model has been trained, it can be deployed to make predictions from data it has never seen before. The accuracy of the results varies depending on the training data and details of the model and training process. Once they have been trained sufficiently, models can be used to automate the tasks of object classification, object detection, pixel classification, image translation, and object tracking. These tasks can be applied to geospatial data obtained from a wide variety of sources including aerial, satellite, drone, motion imagery (video), radar, lidar, bathymetry, feature data, and text.

Application Domains

Some of the fields that can now benefit more easily from the application of deep learning geospatial models using Esri’s pre-trained models include land administration, urban planning, and disaster management.

Pre-trained Models

Pre-trained models can be used to automate the identification and extraction of a wide range of asset classes from satellite, aerial, and drone imagery. Some of the pre-trained models included in Esri’s living atlas include models to extract the following from imagery and point cloud data: building footprints, roads, swimming pools, cars, solar panels, shipwrecks, landcover classifications, human settlements, power lines, and trees. In addition, models are available to track moving vehicles in motion imagery (video) and redact faces and license plates from imagery.